prlq commited on
Commit
c9bb098
·
verified ·
1 Parent(s): 9f32c50

debug: 4 collect-mode issues seen on 536 (nova)

Browse files
debug/536/n012_sawyer_planning.md ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # N012 sawyer_umi: ~47 % success in collect mode, mostly pre_grasp planning failures
2
+
3
+ Reported by the 536 collection machine (agent nova), 2026-11-07. Code: mobile-ready ab32a6159 / 81e58ce6a. Owner said ~40 % is acceptable; this note is for the record.
4
+
5
+ ## Numbers (536, `--collect`)
6
+ | DR | ok | fail |
7
+ |---|---|---|
8
+ | off | 46 | 58 |
9
+ | on | 49 | 48 |
10
+ go2_arm and stretch3 on N012: 49/50 per cell with few failures.
11
+
12
+ ## Failure reasons (last 40 failed sawyer episodes, report `failure_reason`)
13
+ - 16 × `stage pre_grasp failed (planner/skill code 3): failed: global planner ...` (cuRobo MotionGen: "ALL FAILED ... No trajectory satisfied all constraints" although IK solves)
14
+ - 7 × `manipulation: the squeegee is not in the hand after closing`
15
+ - 2 × `violation: squeegee dropped to the floor`; 2 × `push_1 ... target not reached`; 2 × `pre_grasp ... target not reached`; 2 × `grasp ... target not reached`
16
+
17
+ ## Possible lever
18
+ The handoff command sets `MMB_LEAN_SEEDS=1` ("the cuRobo seeds of every N012 Gate 3 run"). We used it for sawyer_umi too. If the Gate 3 sawyer runs did not use the lean seeds, the pre_grasp MotionGen failures (the largest group) may come from it. Worth confirming which seeds the sawyer Gate 3 used.
debug/536/n051_instruction_colour.md ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # N051: collect-mode instruction names the wrong tray colour in ~45 % of successful episodes
2
+
3
+ Reported by the 536 collection machine (amarel xc536, agent nova), 2026-11-07. Code: mobile-ready 6431d1ba1 .. 81e58ce6a (unchanged in this area).
4
+
5
+ ## Symptom
6
+ `scripts/mmb_a11/verify_collect_dataset.py <cell> --task N051 --report <out>/report.json` fails on
7
+ `[FAIL] task text = MobileBench instruction`: the dataset's `meta/tasks.jsonl` says "... pull the cord that brings the **blue** tray to your side",
8
+ while the report of the same episode (`episodes[0]...goal_tray`) says **red** (or the other way round). The trajectory itself is the red one (the task succeeded against goal_tray).
9
+
10
+ 536 counts (successful episodes = ok + verify_fail; failed episodes not counted):
11
+ | robot | DR | kept (instruction correct) | rejected (instruction wrong) |
12
+ |---|---|---|---|
13
+ | go2_arm | off | 11 | 7 |
14
+ | go2_arm | on | 11 | 10 |
15
+ | stretch3 | off | 11 | 8 |
16
+ | stretch3 | on | 11 | 9 |
17
+ 850 saw the same (153 rejected). It does not depend on visual DR or robot: ~45 % of successes, i.e. close to a coin flip on the tray colour.
18
+
19
+ ## Likely cause (from reading the code; not stepped through in a debugger)
20
+ - `collect_mode._n051_slots` returns `{"tray_colour": task_env._ep[env_id]["truth"]["goal_tray"]}` and is called "when the episode is accepted (before the save worker)" (`instructions.set_slot_provider` docstring).
21
+ - `n051_env._sample_truth` draws a new `goal_tray` (random, red/blue) in `_after_episode_reset`.
22
+ - If the episode's reset runs before the recorder accepts the finished trajectory (auto-reset on done, or `finalize_recording` enqueueing after the loop), the provider reads the **next** episode's freshly drawn truth → the colour is right only when the new draw happens to equal the old one (≈50 %), which matches the counts.
23
+ - `report.json` is written from the episode result taken before the reset, so it holds the right colour.
24
+
25
+ ## Suggested fix
26
+ Capture the slot value when the episode starts (or when success is detected) and keep it with the trajectory, instead of reading `task_env._ep` at accept time; or make the provider read the truth stored in the episode result. A test: run 5 seeds with `--collect` and check `tasks.jsonl` colour == report `goal_tray`.
27
+
28
+ ## Impact on the data
29
+ Only episodes whose instruction matches the report were uploaded (the verify step rejects the others), so the HF data is correct; ~45 % of N051 compute is wasted until fixed.
debug/536/n052_stretch3_lift.md ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # N052 stretch3 (and 850's N052 sawyer_umi): every collect episode fails at take_item:lift
2
+
3
+ Reported by the 536 collection machine (agent nova) and 850 (agent data-monster), 2026-11-07. Code: mobile-ready 6431d1ba1 / 7ac467262.
4
+
5
+ ## Symptom
6
+ - N052 × stretch3, `--collect`, DR off and on: **0 / 11** (536). N052 × sawyer_umi on 850: **0 / 13**. N052 go2_arm on the same machines: ~85 % success.
7
+ - Every failure: `reason='skill_failed:take_item:lift:3'` (MemoryStation runner), ~2 min into the episode:
8
+ ```
9
+ event move_target {'label': 'take_item:down:0', 'mode': 4, 'pos': [0.12, 1.1, 0.5136], ...}
10
+ event settle_err {'label': 'take_item:down', 'k': 0, 'err_mm': [1.7, 0.0, -2.9], 'norm_mm': 3.4}
11
+ event move_target {'label': 'take_item:lift', 'mode': 4, 'pos': [0.12, 1.1, 0.6336], ...}
12
+ Env 0 task failed (failed: skill 'take_item:lift' truncated 3)
13
+ ```
14
+ i.e. the hand reaches the item (3.4 mm settle error), closes, and the lift move is truncated (code 3).
15
+
16
+ ## Things worth checking on the build machine
17
+ - The item USD: our machines do not have `/workspace/mobilemanipbench/data/...`; we downloaded `clutter_dice_00000/Object.usd` and `rigid_pen_00010/Object.usd` from the public HF dataset `prlq/dexgrasp-robotwin-4gripper` and bind them at that path. If the build machine's copy differs (scale, collision, mass), grasping with the smaller grippers could fail while go2's works. A checksum of the build machine's `clutter_dice_00000/Object.usd` would settle it.
18
+ - Whether the Gate 3 runs of N052 stretch3 / sawyer used settings that the collect command in DATA_COLLECTION_HANDOFF section 7 does not pass (the MemoryStation command only sets `MEMORY_ROBOT`, `--spawn-jitter 0.3`, `--max-steps 20000`).
debug/536/station_sawyer_collect_crash.md ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # N032 / N033 / N034 sawyer_umi: collect mode crashes on the first step (AttributeError robot_name)
2
+
3
+ Reported by the 536 collection machine (amarel xc536, agent nova), 2026-11-07 22:5x CST.
4
+
5
+ ## Symptom
6
+ - Cells N032 / N033 / N034 × sawyer_umi, `--collect` (visual DR on and off): every episode exits after ~3 min, no dataset.
7
+ - 536 counts: N032 sawyer 0/12, N033 sawyer 0/12, N034 sawyer 0/10. go2_arm and stretch3 cells of the same tasks run normally (N032 go2 4/4, N033 go2 4/4 at the time of writing).
8
+ - Report line: `STATION_RESULT {... "robot": "sawyer_umi", "error": "AttributeError: 'N032EnvXR' object has no attribute 'robot_name'", ...}` (same with `N033EnvXR`, `N034EnvXR`).
9
+
10
+ ## Traceback (identical for the three tasks)
11
+ ```
12
+ File ".../Collect/Task/StationSearch/Env/TestStationEnvXR.py", line 363, in main
13
+ File ".../Task/MobileBench/collect_mode.py", line 423, in after_step
14
+ File ".../Task/MobileBench/collect_mode.py", line 417, in _bind
15
+ self.rob = self.env.scene.robot_manager.robots[self.env.env.robot_name]
16
+ AttributeError: 'N032EnvXR' object has no attribute 'robot_name'
17
+ ```
18
+
19
+ ## Cause
20
+ `collect_mode.HeadTracker` (sawyer_umi only, on by default in collect mode, `MMB_COLLECT_HEAD_TRACK`) looks the robot up lazily on the first `after_step()` through `env.env.robot_name`.
21
+ The MemoryStation runner's env has that attribute (N051–N053 sawyer collect fine), but the StationSearch XR envs (`N032EnvXR`, `N033EnvXR`, `N034EnvXR`, created in `TestStationEnvXR.py`) do not.
22
+ `TestStationEnvXR.py:330` creates the tracker with `HeadTracker.from_env(env, collect_robot)` and line 363 calls `head_track.after_step()`, so every sawyer collect episode dies at the first control step.
23
+ go2_arm and stretch3 are unaffected because `from_env` returns None for robots other than sawyer_umi.
24
+
25
+ ## Suggested fix (for the build machine; we did not change any code)
26
+ Either
27
+ - give `HeadTracker._bind` a fallback when `env.env` has no `robot_name` (e.g. the single entry of `env.scene.robot_manager.robots`, or the XR env's own robot key), or
28
+ - pass the robot asset explicitly from `TestStationEnvXR.py` (`HeadTracker(robot_asset=<sawyer asset>, env=env)`), or
29
+ - add `robot_name` to the StationSearch XR envs.
30
+ A unit test in `tests/station/test_collect_wiring.py` that runs one `after_step()` with `--robot sawyer_umi --collect` would catch it.
31
+
32
+ ## Reproduce
33
+ mobile-ready 81e58ce6a (contains 64eee62be8):
34
+ ```
35
+ OMNI_KIT_ACCEPT_EULA=YES ACCEPT_EULA=Y PYTHONPATH=$PWD/src MAGICSIM_MOTIONGEN_FAILURE_DIAGNOSTICS=1 \
36
+ python -m magicsim.Collect.Task.StationSearch.Env.TestStationEnvXR --robot sawyer_umi --task N032 --level easy --seed 1001 \
37
+ --stage full --episodes 1 --walltime-s 4800 --report OUT/report.json --out OUT/rec \
38
+ --headless --enable_cameras --collect --visual-dr off --set station.dr.enabled=true
39
+ ```
40
+ Workaround that should avoid the crash without a code change (not tried, it changes the head camera): `MMB_COLLECT_HEAD_TRACK=0` (the head then keeps its initial pan).
41
+
42
+ ## Status on 536
43
+ The three sawyer cells were dropped automatically (0 successes in ≥ 8 attempts). They will be collected once a fixed commit is on mobile-ready.