Humanoid LeRobot v2.1 episode samples

This folder contains 40 complete episodes: five representative samples from each of eight logical data sources under /mnt/data/dataset/PretrainData_v21/Humanoid.

The parquet, video, depth and ancillary payloads are byte-for-byte copies of the source files. This is an inspection/interoperability bundle, not a re-indexed drop-in LeRobot dataset: original episode indices are retained in sample.json and manifest.json.

Additional Unitree WBT sample set

unitreerobotics_wbt_10samples/ adds 10 complete episodes from 10 distinct WBT leaf datasets. It covers BrainCo, Dex1, Inspire regular four-view data and all three MainCamOnly two-view variants. This add-on has its own manifest and checksums and is not included in the 40-episode core summary below.

Summary

Source	Robot type	Samples	Frames	Duration (min)	State dim	Action dim	Video files/sample
AgibotWorld2026	g2a	5	5,353	2.97	162, 169, 183, 193	40, 44	6, 7
AgiBot-World-Beta-no-torso-movement_v21	a2d	5	3,544	1.97	55	36	3
HIW-500-LeRobot-v2	unitree_g1	5	3,494	1.94	29	23	3
humanoid-everyday_v21_g1	g1	5	1,949	1.08	69	28	1
humanoid-everyday_v21_h1	h1	5	1,957	1.09	65	26	1
icra_wbc_v21	unitree_g2	5	7,710	4.28	31	44	6
RealSourceData/RealSource-World	RS-02	5	9,553	5.31	71	17	3
unitreerobotics_v21	unitree_g1	5	3,373	1.87	50, 60, 83	50, 60, 97	2, 4

Layout

humanoid_v21_samples/
├── manifest.json          # full bundle manifest
├── manifest.jsonl         # one compact record per episode
├── PAYLOAD_CHECKSUMS.sha256
└── <dataset>/sample_XX/
    ├── sample.json         # task, dimensions, provenance and per-file hashes
    ├── data.parquet        # original episode parquet
    ├── videos/             # registered RGB/depth video streams
    ├── sidecars/           # ancillary parquet or unregistered/overlay depth
    └── source_meta/         # original schema and available source metadata

Selection and edge cases

Samples favor complete 10–60 second episodes and task/schema diversity; one RealSource long-horizon episode is intentionally included.
icra_wbc/sample_* retains head_depth under sidecars/ because that stream exists physically but is absent from info.json features.
agibotworld2026/sample_04 combines the simulation lite RGB/parquet payload with its lite_depth_patch overlay.
humanoid_everyday_g1 and humanoid_everyday_h1 retain their per-episode ancillary parquet files.
Unitree samples intentionally cover the 83/97, 60/60 and 50/50 schemas, including the two-view MainCamOnly variant.
Known missing, shifted, duplicated or malformed source episodes identified during the audit were excluded.

Use manifest.json as the authoritative mapping from every uploaded file to its original path. Verify payload integrity with sha256sum -c PAYLOAD_CHECKSUMS.sha256 from this directory.