Add Gemini failure labels (399 episodes) and prompt documentation

GEMINI_PROMPTS.md

@@ -0,0 +1,287 @@
+# Gemini prompts used for the failure labels
+
+These are the prompt templates exactly as they are stored in `labeler.prompts` of every `labels.json` (identical in all 399 files). `{...}` are Python `str.format` fields filled per episode. Model: `gemini-3.5-flash`; the video is the base-camera `rollout.mp4`.
+
+Notes on how they were applied to this real-robot run:
+
+- The templates were written for a simulator benchmark (LIBERO-X) and reused unchanged, which is why pass 1 says "in LIBERO-X" and mentions a simulator goal predicate. A real run has no simulator predicate: `{failing_predicate}` was filled with the task instruction itself and no simulator detail was available. The model is told the instruction and sees the video; it is not given any ground truth about why the episode failed beyond the human success label (failed episodes only are sent to passes 1-3).
+- The onset is the earliest point at which the mistake responsible for the failure becomes apparent. Pass 1 picks one onset type: `obvious_mistake` (an identifiable task-relevant mistake), `operator_intervention` (no single mistake, but a human operator would step in, e.g. stuck or ineffective repetition) or `timeout` (no earlier identifiable point). It is not the time at which the task becomes unrecoverable.
+- Pipeline: `scripts/semantic_failure` of the labelling project (two-pass onset localiser adapted from dtl184/liberox-evals). With the interactive backend the passes and the keyword phrases run in one shared video session per episode; with the Batch backend each is a separate request on the same uploaded video, with the same prompt text.
+
+## Pass 1 - failure localisation on the full video
+
+_one Gemini chat per episode; the whole rollout video is attached; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_coarse`.
+
+```text
+You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.
+
+The video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.
+
+Task instruction:
+"{task_desc}"
+
+The simulator reports that this goal predicate was NOT satisfied at the end of the episode:
+{failing_predicate}
+
+Additional simulator information:
+{detail}
+
+The simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.
+
+Your job is to determine:
+
+1. the primary visible failure mode,
+2. when the failure behavior begins,
+3. why the episode failed (a long reason AND a short keyword summary),
+4. a recovery action for after the failure has already happened (long AND short),
+5. a prevention action for before the failure has happened (long AND short).
+
+FAILURE ONSET DEFINITION
+
+Use exactly one of the following onset types:
+
+obvious_mistake:
+ Use this when there is an identifiable task-relevant mistake such as
+ dropping an object, knocking an object away, manipulating the wrong
+ object, making a failed placement, or otherwise performing an action
+ that contributes to the eventual failure.
+
+ The onset should be the EARLIEST point where the mistake responsible
+ for the eventual failure becomes apparent.
+
+operator_intervention:
+ Use this when there is not one discrete mistake, but the robot reaches
+ a point where a reasonable human operator would decide that the policy
+ needs assistance.
+
+ Examples include repeatedly attempting an ineffective action, becoming
+ stuck, moving erratically, or clearly ceasing to make useful progress.
+
+timeout:
+ Use this when there is no defensible earlier failure point. If the robot
+ continues behaving plausibly but simply does not finish before the time
+ limit, the failure onset is the END of the rollout.
+
+IMPORTANT RULES
+
+- Do NOT treat the goal predicate being false early in the rollout as failure.
+ Most goal predicates are naturally false until the robot completes the task.
+
+- Do NOT label a temporary mistake as failure onset if the robot later
+ successfully recovers from that mistake.
+
+- Use the earliest mistake or intervention point that actually explains the
+ eventual failed outcome.
+
+- Do not use hindsight to label normal task execution as failure merely because
+ you know the episode eventually fails.
+
+- If there is no identifiable earlier point, use onset_type="timeout" and set
+ failure_onset_seconds to approximately {duration_seconds:.2f}.
+
+FAILURE TAXONOMY
+
+Choose exactly one:
+
+{taxonomy}
+
+REASON, RECOVERY, PREVENTION — TWO VERSIONS EACH
+
+Every language field has a LONG form and a SHORT keyword summary.
+
+LONG: one or two sentences. Specific enough for a human to audit.
+
+SHORT: 2–8 words, semicolon-separated clauses, NOT a full sentence.
+ Same style as the 3-second action phrases. Good for later alignment.
+ Examples: "collision with object; dropping gripped object"
+
+failure_reason / failure_reason_summary:
+ Why this episode failed. Example long: the robot hit the basket while
+ carrying the moka pot, dropped it, and knocked the basket over.
+ Example short: "collision with object; dropping gripped object"
+
+RECOVERY (failure HAS already happened):
+ What the robot should do at or immediately after onset, now that the
+ scene is already broken. Identify object, action, and the condition to
+ restore before continuing.
+ Example long: move the basket away from the hot plate, then place the
+ moka pot on the hot-plate.
+ Example short: "clear basket; place moka on stove"
+
+PREVENTION (failure has NOT yet happened):
+ The counterfactual at or before onset: what the robot should have done
+ so the mistake never occurs. Do not describe post-drop cleanup.
+ Example long: place the moka pot on the stove; do not approach the box.
+ Example short: "place moka on stove; avoid box"
+
+Do not copy taxonomy names into these strings.
+
+Return ONLY a JSON object with this structure:
+
+{{
+ "failure_mode": " ",
+ "onset_type": "obvious_mistake|operator_intervention|timeout",
+ "failure_onset_seconds":,
+ "failure_window_start_seconds":,
+ "failure_window_end_seconds":,
+ "confidence": "high|medium|low",
+ "failure_reason": " ",
+ "failure_reason_summary": " ",
+ "recovery_action": " ",
+ "recovery_action_summary": " ",
+ "prevention_action": " ",
+ "prevention_action_summary": " ",
+ "justification": " "
+}}
+```
+
+## Pass 2 - temporal refinement
+
+_a magnified clip around the coarse onset (frames sampled at 1 FPS, +-3 s) is attached; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_refine`.
+
+```text
+You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.
+
+The complete rollout was already reviewed by another pass of the same VLM.
+
+The coarse analysis identified:
+
+Failure mode:
+{failure_mode}
+
+Failure onset type:
+{onset_type}
+
+Failure reason:
+{failure_reason}
+
+Coarse onset estimate in the original rollout:
+{coarse_seconds:.2f} seconds
+
+Task instruction:
+"{task_desc}"
+
+Failed simulator predicate:
+{failing_predicate}
+
+TEMPORAL MAPPING
+
+This clip has been deliberately slowed down.
+
+Each SECOND of this clip corresponds to exactly ONE FRAME from the original
+rollout.
+
+Therefore:
+
+ refined clip second 0 = original rollout frame {start_frame}
+ refined clip second 1 = original rollout frame {start_frame_plus_one}
+ refined clip second 2 = original rollout frame {start_frame_plus_two}
+ ...
+
+The clip contains {num_frames} original rollout frames.
+
+Your job is ONLY to refine the temporal location of the failure.
+
+Find the EARLIEST frame in this clip where the failure identified above becomes
+visibly apparent or where operator intervention becomes justified.
+
+Remember:
+
+- Do not select a temporary mistake if the robot recovers from it.
+- Do not select a frame merely because the task is not complete yet.
+- Select the first frame associated with the behavior responsible for the
+ eventual failure.
+- If the coarse event is not actually visible in this clip, return null.
+
+Return ONLY:
+
+{{
+ "refined_second":,
+ "confidence": "high|medium|low",
+ "justification": " "
+}}
+```
+
+## Pass 3 - failure language
+
+_same chat; onset, mode and taxonomy are fixed, the model only writes the reason / recovery / prevention text; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_language`.
+
+```text
+You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.
+
+Task instruction:
+"{task_desc}"
+
+Existing labels:
+- failure_mode: {failure_mode}
+- onset_type: {onset_type}
+- failure_onset_seconds: {onset_seconds}
+- failure_reason: {failure_reason}
+- recovery_action: {recovery_action}
+
+Fill the missing language fields.
+
+SHORT STYLE
+2–8 words, semicolon-separated clauses, not a full sentence.
+Example: "collision with object; dropping gripped object"
+
+RECOVERY vs PREVENTION
+- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).
+- prevention_action: long, what the robot should have done at or before onset so the failure never happens.
+- prevention_action_summary: keyword form of that prevention.
+
+Return ONLY:
+
+{{
+ "failure_reason_summary": " ",
+ "recovery_action_summary": " ",
+ "prevention_action": " ",
+ "prevention_action_summary": " "
+}}
+```
+
+## Semantic timeline - 3 second keyword phrases
+
+_same chat; every episode (successes and failures)._  Key in `labels.json`: `labeler.prompts.keyword_phrases`.
+
+```text
+You are reviewing the SAME complete robot manipulation rollout.
+
+The video is {hz:.0f} FPS. Each frame is exactly one executed robot action.
+Episode duration is {duration:.1f} seconds ({n_control} control steps).
+
+Task instruction:
+"{instruction}"
+{failure_context}
+For EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2–6 words)
+for what the manipulator is doing, like "reaching for mug" or "closing gripper"
+or "idle / no progress". Not a full sentence. Only what is visible in THAT window.
+
+Do not copy a failure mode, failure reason, recovery, or prevention into these phrases.
+Those are stored in a separate array. If a window happens to show a mistake,
+describe the visible motion (e.g. "dropping can"), not the taxonomy name.
+
+Windows:
+{windows}
+
+Return ONLY a JSON object:
+{{"segments": [{{"segment_index": 0, "phrase": "..."}}, ...]}}
+Include every segment_index listed above.
+```
+
+## Failure-mode taxonomy (inserted into the pass-1 prompt as {taxonomy})
+
+_the allowed values of `failure_mode`._  Key in `labels.json`: `labeler.prompts.failure_taxonomy`.
+
+```text
+- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.
+- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.
+- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).
+- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.
+- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).
+- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.
+- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).
+- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.
+- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.
+- other: the failure fits none of the labels above. Briefly explain what went wrong.
+```

LABELS.md

@@ -0,0 +1,24 @@
+# Failure labels for the 399 rollouts
+
+Per-episode files next to `rollout.npz` / `rollout.mp4` in each `<task>/<ep>/` folder, plus `manifest.csv` (one row per episode with outcome and onset).
+
+| file | content |
+|---|---|
+| `labels.json` | `success` (human annotation), `semantic_timeline` (3 s keyword phrases per segment with control / policy step ranges), `vlm_failure` (failed episodes: mode, onset time and frame, reason / recovery / prevention text with short summaries, confidence, raw Gemini responses, token usage), `failure_annotation` (onset mapped to `control_step`, `policy_step`, `chunk_index`), `labeler` (model, backend, date, the prompts) |
+| `labels.npz` | the same fields as arrays: `fail_onset_frame`, `fail_onset_seconds`, `fail_mode`, `fail_reason`, `fail_recovery`, `fail_prevention` (+ `_summary`), `sem_t_start`, `sem_t_end`, `sem_control_start`, `sem_control_end`, `sem_phrase` |
+| `example.md` | a readable rendering of the above for that episode |
+
+## Contents
+
+- 399 episodes: 145 successes (no onset), 254 failures with a failure onset.
+- Failure modes (Gemini, one per failed episode): grasp_failure 144, stuck_or_no_progress 31, placement_or_insertion_failure 29, object_displacement 24, open_close_failure 19, timeout_or_insufficient_progress 4, wrong_object 3.
+- Onset types: obvious_mistake 223, operator_intervention 27, timeout 4.
+- Per task (episodes / failures): basket 50/42, blocks 50/44, closer_drawer 50/22, eggplant 50/25, mushroom 49/36, pineapple_out_of_drawer 50/38, push_pineapple 50/23, tomato 50/24.
+
+## Provenance and caveats
+
+- Labeller: Gemini (`gemini-3.5-flash`), prompts in `GEMINI_PROMPTS.md`. 399 episode(s) via `GeminiBackend` (labelled 2026-10-20 to 2026-10-22).
+- The labels are model outputs and have not been individually verified by a person; the success flag comes from the operator. Onset times are the first visible mistake.
+- `vlm_failure_onset_frame` is a control-step index at the recorded control rate (about 50 Hz); the policy step is stored in `failure_annotation.failure_policy_step` and should be used instead of dividing the frame by `replan_steps` (the robot executes about 5.8 control steps per policy call on average, not the nominal 5 or 6).
+- A few onsets are at the very start or very end of a video (`operator_intervention` at the first frames, `timeout` in the last seconds); these are worth a manual look before using them as supervision.
+- Successes carry a keyword-phrase timeline but no failure fields.

basket/ep000/example.md

@@ -0,0 +1,33 @@
+# t002__ep000__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 430  (8.4s at 51 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `basket/ep000/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping basket handle |
+| 3.0–6.0 | 153–305 | 25–50 | pulling handle and losing grip |
+| 6.0–8.4 | 306–429 | 51–71 | retracting gripper and remaining idle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.1434184675834973 s (00:03.1)
+- **frame / control_step:** 160 (policy_step=26, chunk_index=4)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.
+  Summary: lost grasp on handle; basket remained upside down
+- **recovery:** Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.
+  Summary: re-grasp handle securely; flip basket upright
+- **prevention:** Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.
+  Summary: secure firm grasp; rotate handle to flip basket
+- **justification:** None

basket/ep000/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep000__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:43:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket handle",
+      "description": "approaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pulling handle and losing grip",
+      "description": "pulling handle and losing grip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 8.447937131630649,
+      "control_step_start": 306,
+      "control_step_end": 429,
+      "policy_step_start": 51,
+      "policy_step_end": 71,
+      "phrase": "retracting gripper and remaining idle",
+      "description": "retracting gripper and remaining idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.1434184675834973,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 160,
+    "vlm_failure_onset_step": 160,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.2,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.",
+    "vlm_failure_reason_summary": "lost grasp on handle; basket remained upside down",
+    "vlm_recovery_action": "Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp handle securely; flip basket upright",
+    "vlm_prevention_action": "Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.",
+    "vlm_prevention_action_summary": "secure firm grasp; rotate handle to flip basket",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 02:00 (120 seconds), as the robot arm lifts and rotates to flip the basket, the wire handle slips out of the gripper's fingers, causing the basket to fall back onto the table upside down.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 430,
+    "vlm_video_duration_seconds": 8.428290766208251,
+    "vlm_usage": {
+      "input_tokens": 21692,
+      "output_tokens": 325,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.2,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.\",\n \"failure_reason_summary\": \"lost grasp on handle; basket remained upside down\",\n \"recovery_action\": \"Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp handle securely; flip basket upright\",\n \"prevention_action\": \"Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.\",\n \"prevention_action_summary\": \"secure firm grasp; rotate handle to flip basket\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 160,
+    "failure_sim_step": 160,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright."
+  }
+}

basket/ep000/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df71c0f780462ee77b9d4ed9e86399d431d3a2b07da16c4406a4aa0f9767c00b
+size 3839

basket/ep001/example.md

@@ -0,0 +1,33 @@
+# t002__ep001__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 388  (7.8s at 50 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `basket/ep001/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–148 | 0–24 | approaching basket handle |
+| 3.0–6.0 | 149–297 | 24–49 | grasping handle; attempting to flip |
+| 6.0–7.8 | 298–387 | 49–64 | losing grip; retracting upward |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.70281124497992 s (00:05.7)
+- **frame / control_step:** 284 (policy_step=47, chunk_index=2)
+- **coarse onset:** 5.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.
+  Summary: failed handle grasp; gripper slipped off basket
+- **recovery:** Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.
+  Summary: re-align gripper; grasp handle firmly; flip basket upright
+- **prevention:** Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.
+  Summary: center gripper on handle; secure grip before flipping
+- **justification:** At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.

basket/ep001/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep001__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:44:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.991967871485944,
+      "control_step_start": 0,
+      "control_step_end": 148,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "approaching basket handle",
+      "description": "approaching basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.991967871485944,
+      "t_end_sec": 5.983935742971888,
+      "control_step_start": 149,
+      "control_step_end": 297,
+      "policy_step_start": 24,
+      "policy_step_end": 49,
+      "phrase": "grasping handle; attempting to flip",
+      "description": "grasping handle; attempting to flip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.983935742971888,
+      "t_end_sec": 7.791164658634538,
+      "control_step_start": 298,
+      "control_step_end": 387,
+      "policy_step_start": 49,
+      "policy_step_end": 64,
+      "phrase": "losing grip; retracting upward",
+      "description": "losing grip; retracting upward"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.70281124497992,
+    "vlm_failure_onset_timestamp": "00:05.7",
+    "vlm_failure_onset_frame": 284,
+    "vlm_failure_onset_step": 284,
+    "vlm_coarse_failure_onset_seconds": 5.2,
+    "vlm_failure_window_start_seconds": 5.2,
+    "vlm_failure_window_end_seconds": 6.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.",
+    "vlm_failure_reason_summary": "failed handle grasp; gripper slipped off basket",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle firmly; flip basket upright",
+    "vlm_prevention_action": "Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.",
+    "vlm_prevention_action_summary": "center gripper on handle; secure grip before flipping",
+    "vlm_justification": "At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.",
+    "vlm_temporal_justification": "The robot closes its gripper at the handle and begins lifting upward around 02:50. By 02:54 (frame 284 of the rollout), it is visibly apparent that the gripper failed to establish a hold on the handle, lifting away while leaving the basket completely stationary.",
+    "vlm_video_fps": 49.8,
+    "vlm_video_num_frames": 388,
+    "vlm_video_duration_seconds": 7.771084337349398,
+    "vlm_usage": {
+      "input_tokens": 19875,
+      "output_tokens": 394,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.2,\n \"failure_window_start_seconds\": 5.2,\n \"failure_window_end_seconds\": 6.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.\",\n \"failure_reason_summary\": \"failed handle grasp; gripper slipped off basket\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle firmly; flip basket upright\",\n \"prevention_action\": \"Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.\",\n \"prevention_action_summary\": \"center gripper on handle; secure grip before flipping\",\n \"justification\": \"At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 284,
+    "failure_sim_step": 284,
+    "failure_policy_step": 47,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 48,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright."
+  }
+}

basket/ep001/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72be09e4401c14b731b30b53a72ac066426fa90460f605a1b61f0da80789b25e
+size 3841

basket/ep002/example.md

@@ -0,0 +1,33 @@
+# t002__ep002__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 378  (7.4s at 51 Hz)
+- **policy inferences:** 65  (replan_steps=6)
+- **video:** `basket/ep002/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward red basket |
+| 3.0–6.0 | 153–305 | 25–50 | grasping basket; slipping off handle |
+| 6.0–7.4 | 306–377 | 51–62 | hovering arm above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.52054794520548 s (00:04.5)
+- **frame / control_step:** 231 (policy_step=38, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.
+  Summary: losing grip on basket rim; leaving basket tilted
+- **recovery:** Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.
+  Summary: re-grasp basket handle; rotate until fully upright
+- **prevention:** Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.
+  Summary: secure firm grasp on rim; complete rotation fully
+- **justification:** At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.

basket/ep002/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep002__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:46:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward red basket",
+      "description": "reaching toward red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping basket; slipping off handle",
+      "description": "grasping basket; slipping off handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 7.397260273972602,
+      "control_step_start": 306,
+      "control_step_end": 377,
+      "policy_step_start": 51,
+      "policy_step_end": 62,
+      "phrase": "hovering arm above table",
+      "description": "hovering arm above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.52054794520548,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 231,
+    "vlm_failure_onset_step": 231,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 7.38,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.",
+    "vlm_failure_reason_summary": "losing grip on basket rim; leaving basket tilted",
+    "vlm_recovery_action": "Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.",
+    "vlm_recovery_action_summary": "re-grasp basket handle; rotate until fully upright",
+    "vlm_prevention_action": "Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.",
+    "vlm_prevention_action_summary": "secure firm grasp on rim; complete rotation fully",
+    "vlm_justification": "At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.",
+    "vlm_temporal_justification": "At second 180 (03:00 in the video overlay), the robot gripper loses its grip on the basket's handle during the flip attempt, dropping the basket back down onto the surface without orienting it fully upright.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 378,
+    "vlm_video_duration_seconds": 7.3776908023483365,
+    "vlm_usage": {
+      "input_tokens": 21632,
+      "output_tokens": 384,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 7.38,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.\",\n \"failure_reason_summary\": \"losing grip on basket rim; leaving basket tilted\",\n \"recovery_action\": \"Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.\",\n \"recovery_action_summary\": \"re-grasp basket handle; rotate until fully upright\",\n \"prevention_action\": \"Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.\",\n \"prevention_action_summary\": \"secure firm grasp on rim; complete rotation fully\",\n \"justification\": \"At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 231,
+    "failure_sim_step": 231,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table."
+  }
+}

basket/ep002/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac136450fa6768ecf8a5ddd7c15336c388822e6ecd571d77dc7cc508b10237c9
+size 3826

basket/ep003/example.md

@@ -0,0 +1,33 @@
+# t002__ep003__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 311  (6.1s at 51 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `basket/ep003/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and grasping basket edge |
+| 3.0–6.0 | 152–303 | 25–50 | slipping off basket and retracting |
+| 6.0–6.1 | 304–310 | 50–51 | idle hovering above basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.201581027667984 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.
+  Summary: failed grasp on basket edge; slipping off object
+- **recovery:** Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.
+  Summary: re-align gripper; secure basket edge; flip upright
+- **prevention:** Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.
+  Summary: position gripper securely; grasp rim firmly; flip basket
+- **justification:** Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.

basket/ep003/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep003__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:48:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket edge",
+      "description": "approaching and grasping basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping off basket and retracting",
+      "description": "slipping off basket and retracting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.146245059288537,
+      "control_step_start": 304,
+      "control_step_end": 310,
+      "policy_step_start": 50,
+      "policy_step_end": 51,
+      "phrase": "idle hovering above basket",
+      "description": "idle hovering above basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.201581027667984,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.",
+    "vlm_failure_reason_summary": "failed grasp on basket edge; slipping off object",
+    "vlm_recovery_action": "Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; secure basket edge; flip upright",
+    "vlm_prevention_action": "Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.",
+    "vlm_prevention_action_summary": "position gripper securely; grasp rim firmly; flip basket",
+    "vlm_justification": "Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.",
+    "vlm_temporal_justification": "At timestamp 02:42 (frame 162), the robot lowers its gripper onto the top surface of the basket and closes its fingers in an improper position, failing to clamp onto the rim or edge required to flip the basket.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 311,
+    "vlm_video_duration_seconds": 6.126482213438735,
+    "vlm_usage": {
+      "input_tokens": 21439,
+      "output_tokens": 394,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.\",\n \"failure_reason_summary\": \"failed grasp on basket edge; slipping off object\",\n \"recovery_action\": \"Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; secure basket edge; flip upright\",\n \"prevention_action\": \"Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.\",\n \"prevention_action_summary\": \"position gripper securely; grasp rim firmly; flip basket\",\n \"justification\": \"Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright."
+  }
+}

basket/ep003/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07c248f6a33b5a882fcd8a812f016e274fb0f17a9241e02ef0cb32a84deb150d
+size 3832

basket/ep004/example.md

@@ -0,0 +1,34 @@
+# t002__ep004__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 480  (9.4s at 51 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `basket/ep004/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping basket |
+| 3.0–6.0 | 153–305 | 25–50 | lifting and placing basket down |
+| 6.0–9.0 | 306–458 | 51–76 | retracting gripper away |
+| 9.0–9.4 | 459–479 | 76–79 | remaining idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.9411764705882355 s (00:03.9)
+- **frame / control_step:** 201 (policy_step=33, chunk_index=3)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.
+  Summary: failed to rotate basket; incorrect placement orientation
+- **recovery:** Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.
+  Summary: re-grasp basket; rotate upright; place flat on table
+- **prevention:** Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.
+  Summary: rotate gripper to flip basket upright; verify orientation before releasing
+- **justification:** None

basket/ep004/labels.json

@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep004__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:31:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket",
+      "description": "approaching and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and placing basket down",
+      "description": "lifting and placing basket down"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "retracting gripper away",
+      "description": "retracting gripper away"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 9.411764705882353,
+      "control_step_start": 459,
+      "control_step_end": 479,
+      "policy_step_start": 76,
+      "policy_step_end": 79,
+      "phrase": "remaining idle above table",
+      "description": "remaining idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.9411764705882355,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 201,
+    "vlm_failure_onset_step": 201,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.",
+    "vlm_failure_reason_summary": "failed to rotate basket; incorrect placement orientation",
+    "vlm_recovery_action": "Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.",
+    "vlm_recovery_action_summary": "re-grasp basket; rotate upright; place flat on table",
+    "vlm_prevention_action": "Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.",
+    "vlm_prevention_action_summary": "rotate gripper to flip basket upright; verify orientation before releasing",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot lifts the basket by its handle to peak height, but at second 160 it begins descending back towards the table surface without attempting to rotate or flip the basket into an upright orientation.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 480,
+    "vlm_video_duration_seconds": 9.392156862745098,
+    "vlm_usage": {
+      "input_tokens": 21755,
+      "output_tokens": 321,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.\",\n \"failure_reason_summary\": \"failed to rotate basket; incorrect placement orientation\",\n \"recovery_action\": \"Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.\",\n \"recovery_action_summary\": \"re-grasp basket; rotate upright; place flat on table\",\n \"prevention_action\": \"Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.\",\n \"prevention_action_summary\": \"rotate gripper to flip basket upright; verify orientation before releasing\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 201,
+    "failure_sim_step": 201,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position."
+  }
+}

basket/ep004/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf3cb55a2743d032de650f7be061ddc1957a1450fe4b5afe5ae5e55fa9f918de
+size 3911

basket/ep005/example.md

@@ -0,0 +1,33 @@
+# t002__ep005__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 334  (6.7s at 50 Hz)
+- **policy inferences:** 58  (replan_steps=6)
+- **video:** `basket/ep005/rollout.mp4`
+- **features:** `58` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching basket and slipping |
+| 3.0–6.0 | 151–301 | 25–50 | attempting second grasp on basket |
+| 6.0–6.7 | 302–333 | 50–55 | retracting end effector upward |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.549800796812749 s (00:02.5)
+- **frame / control_step:** 128 (policy_step=21, chunk_index=2)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.
+  Summary: failed grasp on basket; slipped off handle
+- **recovery:** Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.
+  Summary: reposition gripper; secure grip on edge; flip basket upright
+- **prevention:** Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.
+  Summary: align gripper with rim; secure firm grip; rotate to flip
+- **justification:** At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.

basket/ep005/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep005__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:31:51+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching basket and slipping",
+      "description": "approaching basket and slipping"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting second grasp on basket",
+      "description": "attempting second grasp on basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.653386454183266,
+      "control_step_start": 302,
+      "control_step_end": 333,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "retracting end effector upward",
+      "description": "retracting end effector upward"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.549800796812749,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 128,
+    "vlm_failure_onset_step": 128,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 1.8,
+    "vlm_failure_window_end_seconds": 2.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.",
+    "vlm_failure_reason_summary": "failed grasp on basket; slipped off handle",
+    "vlm_recovery_action": "Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.",
+    "vlm_recovery_action_summary": "reposition gripper; secure grip on edge; flip basket upright",
+    "vlm_prevention_action": "Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.",
+    "vlm_prevention_action_summary": "align gripper with rim; secure firm grip; rotate to flip",
+    "vlm_justification": "At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.",
+    "vlm_temporal_justification": "At 02:08 (128 seconds), as the robotic arm pulls upward to flip the basket, the gripper fingers lose their hold on the handle and slip off completely, leaving the basket in its upside-down position.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 334,
+    "vlm_video_duration_seconds": 6.633466135458167,
+    "vlm_usage": {
+      "input_tokens": 18166,
+      "output_tokens": 389,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 1.8,\n \"failure_window_end_seconds\": 2.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.\",\n \"failure_reason_summary\": \"failed grasp on basket; slipped off handle\",\n \"recovery_action\": \"Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.\",\n \"recovery_action_summary\": \"reposition gripper; secure grip on edge; flip basket upright\",\n \"prevention_action\": \"Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.\",\n \"prevention_action_summary\": \"align gripper with rim; secure firm grip; rotate to flip\",\n \"justification\": \"At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 128,
+    "failure_sim_step": 128,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright."
+  }
+}

basket/ep005/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a04c4bba2c259a6926ac1140c360f295c1148d5da2301d09134a66643a13d2b5
+size 3880

basket/ep006/example.md

@@ -0,0 +1,33 @@
+# t002__ep006__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 328  (6.5s at 50 Hz)
+- **policy inferences:** 57  (replan_steps=5)
+- **video:** `basket/ep006/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | attempting grasp on basket handle |
+| 3.0–6.0 | 150–299 | 30–59 | repeating grasp attempt on handle |
+| 6.0–6.5 | 300–327 | 60–65 | hovering above red basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.8962075848303392 s (00:01.9)
+- **frame / control_step:** 95 (policy_step=19, chunk_index=0)
+- **coarse onset:** 1.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.
+  Summary: failed handle grasp; missing object grip
+- **recovery:** Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.
+  Summary: realign with handle; grasp securely; flip basket upright
+- **prevention:** Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.
+  Summary: align gripper with handle; grasp firmly; flip basket upright
+- **justification:** At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.

basket/ep006/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep006__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:32:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "attempting grasp on basket handle",
+      "description": "attempting grasp on basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "repeating grasp attempt on handle",
+      "description": "repeating grasp attempt on handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.54690618762475,
+      "control_step_start": 300,
+      "control_step_end": 327,
+      "policy_step_start": 60,
+      "policy_step_end": 65,
+      "phrase": "hovering above red basket",
+      "description": "hovering above red basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.8962075848303392,
+    "vlm_failure_onset_timestamp": "00:01.9",
+    "vlm_failure_onset_frame": 95,
+    "vlm_failure_onset_step": 95,
+    "vlm_coarse_failure_onset_seconds": 1.5,
+    "vlm_failure_window_start_seconds": 1.0,
+    "vlm_failure_window_end_seconds": 2.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.",
+    "vlm_failure_reason_summary": "failed handle grasp; missing object grip",
+    "vlm_recovery_action": "Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.",
+    "vlm_recovery_action_summary": "realign with handle; grasp securely; flip basket upright",
+    "vlm_prevention_action": "Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.",
+    "vlm_prevention_action_summary": "align gripper with handle; grasp firmly; flip basket upright",
+    "vlm_justification": "At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.",
+    "vlm_temporal_justification": "At timestamp 01:35 (95 seconds into the clip), the robot gripper fingers close completely without pinching or securing the basket's black handle, making the grasp failure visibly obvious before the arm lifts away empty.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 328,
+    "vlm_video_duration_seconds": 6.526946107784431,
+    "vlm_usage": {
+      "input_tokens": 16521,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 1.5,\n \"failure_window_start_seconds\": 1.0,\n \"failure_window_end_seconds\": 2.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.\",\n \"failure_reason_summary\": \"failed handle grasp; missing object grip\",\n \"recovery_action\": \"Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.\",\n \"recovery_action_summary\": \"realign with handle; grasp securely; flip basket upright\",\n \"prevention_action\": \"Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.\",\n \"prevention_action_summary\": \"align gripper with handle; grasp firmly; flip basket upright\",\n \"justification\": \"At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 95,
+    "failure_sim_step": 95,
+    "failure_policy_step": 19,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 20,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position."
+  }
+}

basket/ep006/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b3fc4d7c3dca92a591ba079bc18fca46723e01bd83122cce08183069198850
+size 3847

basket/ep007/example.md

@@ -0,0 +1,32 @@
+# t002__ep007__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 286  (5.6s at 51 Hz)
+- **policy inferences:** 50  (replan_steps=6)
+- **video:** `basket/ep007/rollout.mp4`
+- **features:** `50` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and gripping handle |
+| 3.0–5.6 | 152–285 | 25–47 | losing grip and retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.662721893491124 s (00:02.7)
+- **frame / control_step:** 135 (policy_step=22, chunk_index=3)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.
+  Summary: insecure grip on handle; handle slipping during lift
+- **recovery:** Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.
+  Summary: re-grasp handle securely; rotate and flip basket upright
+- **prevention:** Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.
+  Summary: align gripper deeper on handle; establish solid grip before flipping
+- **justification:** At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.

basket/ep007/labels.json

@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep007__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:32:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and gripping handle",
+      "description": "reaching for and gripping handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.6410256410256405,
+      "control_step_start": 152,
+      "control_step_end": 285,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "losing grip and retracting arm",
+      "description": "losing grip and retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.662721893491124,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 135,
+    "vlm_failure_onset_step": 135,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.",
+    "vlm_failure_reason_summary": "insecure grip on handle; handle slipping during lift",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp handle securely; rotate and flip basket upright",
+    "vlm_prevention_action": "Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.",
+    "vlm_prevention_action_summary": "align gripper deeper on handle; establish solid grip before flipping",
+    "vlm_justification": "At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.",
+    "vlm_temporal_justification": "At clip second 02:15 (135 seconds), the robot begins lifting its arm after closing the gripper, but fails to maintain a secure grasp on the basket handle, causing the handle to slip out as the arm raises.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 286,
+    "vlm_video_duration_seconds": 5.621301775147929,
+    "vlm_usage": {
+      "input_tokens": 19863,
+      "output_tokens": 386,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.\",\n \"failure_reason_summary\": \"insecure grip on handle; handle slipping during lift\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp handle securely; rotate and flip basket upright\",\n \"prevention_action\": \"Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.\",\n \"prevention_action_summary\": \"align gripper deeper on handle; establish solid grip before flipping\",\n \"justification\": \"At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 135,
+    "failure_sim_step": 135,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright."
+  }
+}

basket/ep007/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7458f115493159d30b77abda64bca148a6446b55c90072b57d3805ec73458ef
+size 3765

basket/ep008/example.md

@@ -0,0 +1,22 @@
+# t002__ep008__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 522  (10.3s at 51 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `basket/ep008/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching the red basket |
+| 3.0–6.0 | 152–303 | 25–50 | grasping the basket handle |
+| 6.0–9.0 | 304–455 | 50–75 | flipping the basket upright |
+| 9.0–10.3 | 456–521 | 76–86 | releasing the basket handle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.

basket/ep008/labels.json

@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t002__ep008__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:33:02+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the red basket",
+      "description": "approaching the red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the basket handle",
+      "description": "grasping the basket handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "flipping the basket upright",
+      "description": "flipping the basket upright"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 10.29585798816568,
+      "control_step_start": 456,
+      "control_step_end": 521,
+      "policy_step_start": 76,
+      "policy_step_end": 86,
+      "phrase": "releasing the basket handle",
+      "description": "releasing the basket handle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}

basket/ep008/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc69e0c11b35a4c0e02d6a1c26c9a8acbfd81906d012e13c902d2a879ec5272a
+size 3106

basket/ep009/example.md

@@ -0,0 +1,32 @@
+# t002__ep009__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 275  (5.4s at 51 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `basket/ep009/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | grasping basket handle and lifting |
+| 3.0–5.4 | 152–274 | 25–45 | dropping basket and raising gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.102766798418972 s (00:03.1)
+- **frame / control_step:** 157 (policy_step=26, chunk_index=1)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.
+  Summary: dropping gripped basket; failed to flip upright
+- **recovery:** Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.
+  Summary: regrasp basket handle; flip upright
+- **prevention:** Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.
+  Summary: maintain firm grip; rotate basket carefully
+- **justification:** At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.

basket/ep009/labels.json

@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep009__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:33:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "grasping basket handle and lifting",
+      "description": "grasping basket handle and lifting"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 5.434782608695652,
+      "control_step_start": 152,
+      "control_step_end": 274,
+      "policy_step_start": 25,
+      "policy_step_end": 45,
+      "phrase": "dropping basket and raising gripper",
+      "description": "dropping basket and raising gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.102766798418972,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 157,
+    "vlm_failure_onset_step": 157,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.8,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.",
+    "vlm_failure_reason_summary": "dropping gripped basket; failed to flip upright",
+    "vlm_recovery_action": "Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.",
+    "vlm_recovery_action_summary": "regrasp basket handle; flip upright",
+    "vlm_prevention_action": "Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.",
+    "vlm_prevention_action_summary": "maintain firm grip; rotate basket carefully",
+    "vlm_justification": "At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.",
+    "vlm_temporal_justification": "At timestamp 02:37 (157 seconds into the clip), the robot gripper loses its grip on the handle of the red basket, causing it to slip out and drop back onto the table.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 275,
+    "vlm_video_duration_seconds": 5.41501976284585,
+    "vlm_usage": {
+      "input_tokens": 19488,
+      "output_tokens": 359,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.8,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.\",\n \"failure_reason_summary\": \"dropping gripped basket; failed to flip upright\",\n \"recovery_action\": \"Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.\",\n \"recovery_action_summary\": \"regrasp basket handle; flip upright\",\n \"prevention_action\": \"Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.\",\n \"prevention_action_summary\": \"maintain firm grip; rotate basket carefully\",\n \"justification\": \"At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 157,
+    "failure_sim_step": 157,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "object_displacement",
+    "correction_action": "Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table."
+  }
+}

basket/ep009/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9ed86b6765379af746a157b9e3e96adc20a2a19eff5c3a319649f0576717f43
+size 3745

basket/ep010/example.md

@@ -0,0 +1,33 @@
+# t002__ep010__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 320  (6.3s at 51 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `basket/ep010/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | positioning gripper above basket |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper and bumping handle |
+| 6.0–6.3 | 304–319 | 50–53 | remaining near displaced basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.803149606299213 s (00:04.8)
+- **frame / control_step:** 244 (policy_step=40, chunk_index=4)
+- **coarse onset:** 4.9 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.
+  Summary: collision with handle; object displacement
+- **recovery:** Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.
+  Summary: realign gripper; grasp handle; flip basket upright
+- **prevention:** Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.
+  Summary: align over handle; lower precisely; execute grasp
+- **justification:** At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.

basket/ep010/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep010__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:34:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "positioning gripper above basket",
+      "description": "positioning gripper above basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper and bumping handle",
+      "description": "lowering gripper and bumping handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 6.299212598425197,
+      "control_step_start": 304,
+      "control_step_end": 319,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "remaining near displaced basket",
+      "description": "remaining near displaced basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.803149606299213,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 244,
+    "vlm_failure_onset_step": 244,
+    "vlm_coarse_failure_onset_seconds": 4.9,
+    "vlm_failure_window_start_seconds": 4.4,
+    "vlm_failure_window_end_seconds": 5.4,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.",
+    "vlm_failure_reason_summary": "collision with handle; object displacement",
+    "vlm_recovery_action": "Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.",
+    "vlm_recovery_action_summary": "realign gripper; grasp handle; flip basket upright",
+    "vlm_prevention_action": "Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.",
+    "vlm_prevention_action_summary": "align over handle; lower precisely; execute grasp",
+    "vlm_justification": "At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.",
+    "vlm_temporal_justification": "At clip second 02:27 (frame 147), the robot end-effector makes direct contact with the top handle of the basket due to misalignment, causing the object to displace rather than being cleanly grasped.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 320,
+    "vlm_video_duration_seconds": 6.2795275590551185,
+    "vlm_usage": {
+      "input_tokens": 16277,
+      "output_tokens": 372,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.9,\n \"failure_window_start_seconds\": 4.4,\n \"failure_window_end_seconds\": 5.4,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.\",\n \"failure_reason_summary\": \"collision with handle; object displacement\",\n \"recovery_action\": \"Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.\",\n \"recovery_action_summary\": \"realign gripper; grasp handle; flip basket upright\",\n \"prevention_action\": \"Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.\",\n \"prevention_action_summary\": \"align over handle; lower precisely; execute grasp\",\n \"justification\": \"At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 244,
+    "failure_sim_step": 244,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "object_displacement",
+    "correction_action": "Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright."
+  }
+}

basket/ep010/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c6dbd6413d720b66681a424fd950b7814c841c015e0f0932aced31536520508
+size 3818

basket/ep011/example.md

@@ -0,0 +1,33 @@
+# t002__ep011__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 330  (6.5s at 51 Hz)
+- **policy inferences:** 56  (replan_steps=6)
+- **video:** `basket/ep011/rollout.mp4`
+- **features:** `56` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | losing grip and dropping basket |
+| 6.0–6.5 | 304–329 | 50–54 | retracting gripper away |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.201581027667984 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 3.2 s (refined=False)
+- **mode:** object_displacement
+- **reason:** The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.
+  Summary: lost grip while flipping; basket dropped improperly
+- **recovery:** Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.
+  Summary: re-grasp handle; rotate basket upright
+- **prevention:** Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.
+  Summary: secure handle grip; rotate smoothly until upright
+- **justification:** Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.

basket/ep011/labels.json

@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep011__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:34:38+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "losing grip and dropping basket",
+      "description": "losing grip and dropping basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.521739130434782,
+      "control_step_start": 304,
+      "control_step_end": 329,
+      "policy_step_start": 50,
+      "policy_step_end": 54,
+      "phrase": "retracting gripper away",
+      "description": "retracting gripper away"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.201581027667984,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.8,
+    "vlm_failure_window_end_seconds": 3.8,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.",
+    "vlm_failure_reason_summary": "lost grip while flipping; basket dropped improperly",
+    "vlm_recovery_action": "Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.",
+    "vlm_recovery_action_summary": "re-grasp handle; rotate basket upright",
+    "vlm_prevention_action": "Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.",
+    "vlm_prevention_action_summary": "secure handle grip; rotate smoothly until upright",
+    "vlm_justification": "Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.",
+    "vlm_temporal_justification": "At 02:24, the robot gripper slips off the black handle while trying to lift and flip the basket, causing the basket to drop back onto the table upside down.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 330,
+    "vlm_video_duration_seconds": 6.5019762845849804,
+    "vlm_usage": {
+      "input_tokens": 21506,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.8,\n \"failure_window_end_seconds\": 3.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.\",\n \"failure_reason_summary\": \"lost grip while flipping; basket dropped improperly\",\n \"recovery_action\": \"Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.\",\n \"recovery_action_summary\": \"re-grasp handle; rotate basket upright\",\n \"prevention_action\": \"Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.\",\n \"prevention_action_summary\": \"secure handle grip; rotate smoothly until upright\",\n \"justification\": \"Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface."
+  }
+}

basket/ep011/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:057eaba6256318f7f850886822db572dbeeec54d7d8316e99761f70350f3a09a
+size 3816

basket/ep012/example.md

@@ -0,0 +1,32 @@
+# t002__ep012__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 249  (4.9s at 51 Hz)
+- **policy inferences:** 43  (replan_steps=6)
+- **video:** `basket/ep012/rollout.mp4`
+- **features:** `43` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | attempting grasp on basket handle |
+| 3.0–4.9 | 153–248 | 25–41 | retracting end-effector and hovering |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.8003913894324852 s (00:01.8)
+- **frame / control_step:** 92 (policy_step=15, chunk_index=2)
+- **coarse onset:** 1.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.
+  Summary: failed grasp attempt; missing basket handle
+- **recovery:** Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.
+  Summary: re-align gripper; grasp handle; flip basket upright
+- **prevention:** Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.
+  Summary: align gripper around handle; ensure secure grasp
+- **justification:** At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.

basket/ep012/labels.json

@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep012__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "attempting grasp on basket handle",
+      "description": "attempting grasp on basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 4.87279843444227,
+      "control_step_start": 153,
+      "control_step_end": 248,
+      "policy_step_start": 25,
+      "policy_step_end": 41,
+      "phrase": "retracting end-effector and hovering",
+      "description": "retracting end-effector and hovering"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.8003913894324852,
+    "vlm_failure_onset_timestamp": "00:01.8",
+    "vlm_failure_onset_frame": 92,
+    "vlm_failure_onset_step": 92,
+    "vlm_coarse_failure_onset_seconds": 1.2,
+    "vlm_failure_window_start_seconds": 1.0,
+    "vlm_failure_window_end_seconds": 2.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.",
+    "vlm_failure_reason_summary": "failed grasp attempt; missing basket handle",
+    "vlm_recovery_action": "Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle; flip basket upright",
+    "vlm_prevention_action": "Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.",
+    "vlm_prevention_action_summary": "align gripper around handle; ensure secure grasp",
+    "vlm_justification": "At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.",
+    "vlm_temporal_justification": "At second 92 (01:32), the robot begins closing its gripper prematurely above the basket's handle loop without making contact or fully lowering around the handle, failing to secure the object.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 249,
+    "vlm_video_duration_seconds": 4.853228962818004,
+    "vlm_usage": {
+      "input_tokens": 15702,
+      "output_tokens": 357,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 1.2,\n \"failure_window_start_seconds\": 1.0,\n \"failure_window_end_seconds\": 2.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.\",\n \"failure_reason_summary\": \"failed grasp attempt; missing basket handle\",\n \"recovery_action\": \"Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle; flip basket upright\",\n \"prevention_action\": \"Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.\",\n \"prevention_action_summary\": \"align gripper around handle; ensure secure grasp\",\n \"justification\": \"At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 92,
+    "failure_sim_step": 92,
+    "failure_policy_step": 15,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 16,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright."
+  }
+}

basket/ep012/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8ce1199248b976f93a9fab382603c3e254cb3add8a53d09547f1e2d138130f0
+size 3772

basket/ep013/example.md

@@ -0,0 +1,21 @@
+# t002__ep013__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 323  (6.4s at 50 Hz)
+- **policy inferences:** 57  (replan_steps=6)
+- **video:** `basket/ep013/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and grasping basket handle |
+| 3.0–6.0 | 150–299 | 25–49 | flipping basket upright |
+| 6.0–6.4 | 300–322 | 50–53 | retracting arm away |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.

basket/ep013/labels.json

@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t002__ep013__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "flipping basket upright",
+      "description": "flipping basket upright"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.447105788423153,
+      "control_step_start": 300,
+      "control_step_end": 322,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}

basket/ep013/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07498299c7bba88eb16556d34a7d4f55e4cb0e2783fa728cc3f42e1bf9f77ea8
+size 3081

basket/ep014/example.md

@@ -0,0 +1,32 @@
+# t002__ep014__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 274  (5.4s at 50 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `basket/ep014/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping basket |
+| 3.0–5.4 | 151–273 | 25–45 | lifting inverted basket straight up |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.5317460317460316 s (00:03.5)
+- **frame / control_step:** 178 (policy_step=29, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.
+  Summary: lifting basket straight up; failing to rotate upright
+- **recovery:** Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.
+  Summary: lower basket; rotate wrist; set upright on table
+- **prevention:** Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.
+  Summary: rotate end-effector while lifting; turn basket right-side up
+- **justification:** Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.

basket/ep014/labels.json

@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep014__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping basket",
+      "description": "reaching for and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.436507936507937,
+      "control_step_start": 151,
+      "control_step_end": 273,
+      "policy_step_start": 25,
+      "policy_step_end": 45,
+      "phrase": "lifting inverted basket straight up",
+      "description": "lifting inverted basket straight up"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.5317460317460316,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 178,
+    "vlm_failure_onset_step": 178,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.",
+    "vlm_failure_reason_summary": "lifting basket straight up; failing to rotate upright",
+    "vlm_recovery_action": "Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.",
+    "vlm_recovery_action_summary": "lower basket; rotate wrist; set upright on table",
+    "vlm_prevention_action": "Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.",
+    "vlm_prevention_action_summary": "rotate end-effector while lifting; turn basket right-side up",
+    "vlm_justification": "Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.",
+    "vlm_temporal_justification": "At timestamp 02:58 (frame 178), after grasping the strap, the robot arm begins lifting the basket straight upward off the table without attempting any flipping or rotation maneuver, causing the basket to remain upside down in mid-air.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 274,
+    "vlm_video_duration_seconds": 5.416666666666667,
+    "vlm_usage": {
+      "input_tokens": 19426,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.\",\n \"failure_reason_summary\": \"lifting basket straight up; failing to rotate upright\",\n \"recovery_action\": \"Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.\",\n \"recovery_action_summary\": \"lower basket; rotate wrist; set upright on table\",\n \"prevention_action\": \"Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.\",\n \"prevention_action_summary\": \"rotate end-effector while lifting; turn basket right-side up\",\n \"justification\": \"Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 178,
+    "failure_sim_step": 178,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down."
+  }
+}

basket/ep014/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:611f1a62f4d46653ea607835ceb7855ebcd6eec804b83c15b09ce13d398b09c9
+size 3842

basket/ep015/example.md

@@ -0,0 +1,34 @@
+# t002__ep015__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 463  (9.1s at 51 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `basket/ep015/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and missing basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | retracting and hovering over basket |
+| 6.0–9.0 | 304–455 | 50–75 | hovering near basket |
+| 9.0–9.1 | 456–462 | 76–77 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.677165354330709 s (00:02.7)
+- **frame / control_step:** 136 (policy_step=22, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.
+  Summary: failed grasp on basket handle; empty retraction
+- **recovery:** Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.
+  Summary: re-grasp basket handle; flip basket upright
+- **prevention:** Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.
+  Summary: align gripper with handle; secure grip before flipping
+- **justification:** At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.

basket/ep015/labels.json

@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep015__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:36:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and missing basket handle",
+      "description": "approaching and missing basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting and hovering over basket",
+      "description": "retracting and hovering over basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering near basket",
+      "description": "hovering near basket"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.114173228346457,
+      "control_step_start": 456,
+      "control_step_end": 462,
+      "policy_step_start": 76,
+      "policy_step_end": 77,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.677165354330709,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 136,
+    "vlm_failure_onset_step": 136,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; empty retraction",
+    "vlm_recovery_action": "Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp basket handle; flip basket upright",
+    "vlm_prevention_action": "Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.",
+    "vlm_prevention_action_summary": "align gripper with handle; secure grip before flipping",
+    "vlm_justification": "At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.",
+    "vlm_temporal_justification": "At timestamp 02:16 (frame 136), the robot gripper finishes closing without properly capturing or securing the wire handle of the basket. As the arm lifts shortly after, the handle remains on the table, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 463,
+    "vlm_video_duration_seconds": 9.094488188976378,
+    "vlm_usage": {
+      "input_tokens": 21627,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; empty retraction\",\n \"recovery_action\": \"Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp basket handle; flip basket upright\",\n \"prevention_action\": \"Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.\",\n \"prevention_action_summary\": \"align gripper with handle; secure grip before flipping\",\n \"justification\": \"At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 136,
+    "failure_sim_step": 136,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright."
+  }
+}

basket/ep015/labels.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e988cb64cfb79fef1ba3c6d9c82f20fd82972730f5ff3a9b52bf0be9374e7c24
+size 3829