LABELS.md

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.