yam7h-ego

Retargeted egocentric human video, YAM bimanual.

Part of a 7.014 h co-fine-tuning mixture for pi0-FAST on the YAM bimanual arm:

source	episodes	frames	hours
angkul07/yam7h-teleop	761	251,703	2.331
angkul07/yam7h-ego	4,174	505,837	4.684
total	4,935	757,540	7.014

This dataset

Episodes	4,174
Frames	505,837
Duration	4.684 h
FPS	30
Codebase version	v2.1
Source	angkul07/EgoDex-PickPlace-YAM-14dof-multiview
Views	observation.images.top, .left_wrist, .right_wrist (224x224)
State / action	14-D (2x6 arm joints + 2 grippers)

Selection

Retargeted egocentric human video (EgoDex -> YAM 14-DoF), filtered and subsampled.

Constraint: at least 30 episodes per object. Objects are not a metadata field — they are parsed from the task sentence (Pick up a <object> from the <source> and place it ...). The parser is validated against the published category table and reproduces all 269 distinct objects across 9 categories exactly; select_mixture.py hard-fails if that check ever stops matching.

stage	objects	episodes	frames	hours
source	269	8,842	1,074,893	9.953
min 30 eps/object	111	6,548	794,288	7.355
subsampled to target	111	4,174	505,837	4.684

The threshold is applied to the source pool; the subsample is then proportional per object, so all 111 objects survive and the object mix is preserved. Note the smallest object ends at 20 episodes in the delivered data, since it keeps its proportional share.

471 task strings are retained verbatim in tasks.jsonl so task_index values stay valid; not all are used after subsampling.

Known domain gaps vs teleop

Measured, not corrected — relevant if you train on this mixture:

• Gripper range collision. Teleop's right gripper is crisply bimodal (closed 0.333 / open 0.984, 3.16 sigma). Retargeted ego is compressed (0.222 / 0.429, 1.83 sigma), so ego's open lands on teleop's closed. Both use 0_closed_1_open; there is no polarity inversion, but the ranges overlap.
• Synthesized wrist views are soft. Laplacian-variance sharpness, teleop top/L/R = 1043.7 / 420.6 / 421.8 versus ego 93.4 / 62.2 / 38.1. This is missing source information from edge-clamped warps, not compression — ego top and right_wrist have near-identical bitrate yet differ 2.7x in sharpness.
• No corrective signal. action[t] == state[t+1] exactly for ego, versus a 0.683 correlation for teleop. Ego joint deltas are 2-4x narrower with clipping pile-ups at exactly +/-0.1 and +/-0.2.

A per-episode retargeting QA sidecar travels with the source dataset; the active_ik_mean_cm <= 15 filter it recommends was not applied here (3,948 of the 4,174 selected episodes would pass it).

Episode renumbering

Episodes are renumbered 0..4173; original source indices are in source_episodes.json (episodes list + index_map).

This is required, not cosmetic: lerobot's get_episodes_file_paths probes range(meta.total_episodes), so a folder holding sparse original indices is unopenable — it silently falls back to the Hub and 401s.

Verification

Built by vast_run/build_mixture.py and checked by vast_run/audit_mixture.py in angkul07/openpi:

• 25 random episodes read back and compared column-by-column against the source parquet — observation.state, action, timestamp, frame_index, task_index all byte-identical. Renumbering touches only episode_index and the global index.
• 8 episodes x 3 cameras decoded with ffprobe; every frame count matches its parquet row count.
• Episode indices contiguous, global index contiguous across the dataset, all parquet and video files present.

Reproducing the selection

vast_run/yam7h_manifest.json records the exact episode lists (seed 0).

python vast_run/select_mixture.py --dry-run     # plan, downloads only meta/
python vast_run/build_mixture.py                # fetch only selected episodes
python vast_run/audit_mixture.py                # independent verification