Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
label
class label
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
1observation_images_left_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
2observation_images_right_wrist
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
0observation_images_head
End of preview. Expand in Data Studio

egg_cake_cleaned — 蛋烘糕双臂机器人轨迹(自动清洗版)

Source: C0notSilly/egg_cake (3 batches, 216 episodes). This repo contains the 197 episodes that passed automatic cleaning, with skill boundaries repaired. LeRobot v2 layout per batch (data/, videos/, meta/), episodes renumbered 0..N-1; meta/episode_mapping.jsonl maps new → source index.

batch source eps kept eps with repaired boundaries frames videos
dataset_2026-09-10_14-32-06 31 29 4 41891 85
dataset_2026-09-12_12-01-37 124 111 111 (machine re-segmentation) 187961 325
dataset_2026-09-15_18-00-19 61 57 21 90239 133

Batch 09-12 note: its original skill boundaries were too noisy to repair boundary-by-boundary (waiting periods assigned inconsistently), so this batch was re-segmented by machine: skill order kept from the annotation, boundaries re-estimated by DTW against clean-batch (09-15) skill templates and placed at the inter-action pause (end-effector speed minimum). Original breakpoints are kept in ; per-episode GPT-5.5 boundary-check counts are in cleaning.vlm_boundary_check (informational — the checker cannot judge boundaries inside static waiting periods).

Cleaning pipeline (blind, no human labels)

  1. Segmentation check & repair — per-skill trajectory clustering (ee pose + gripper) → DTW alignment of each episode to a template built in its own skill order → boundaries disagreeing with the annotation by >2 s are arbitrated by GPT-5.5 on anonymised frame strips (A/B, source hidden). Boundaries where the judge favoured the DTW estimate are moved there (refined to the local end-effector speed minimum, ±0.5 s) and skill_id is rewritten per frame; meta/episodes.jsonl breakpoints updated. Episodes with an undecidable boundary, a move crossing its neighbour, or ≤2 skills (aborted) are dropped (meta/dropped_episodes.json; full detail in meta/repair_log.json).
  2. Video screening (GPT-5.5) — pan-state based task-completion check + scene-anomaly check; informational in this release (aborted episodes are the only video-derived drops).

Known upstream issues carried over: some episodes lack one or more camera videos (see total_videos); batch 09-12 annotations were the noisiest (re-segmented, 111/124 kept).

Generated 2026-09-17 by the auto-screen pipeline (report: internal Feishu doc unified_aligned_v2 两层自动清洗结果).

Downloads last month
94