README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - robotics
tags:
  - lerobot
  - bigym
  - mujoco
  - 3d-gaussian-splatting
  - imitation-learning
pretty_name: Canonical BiGym 3DGS Cutlery 32 A800 Collection
extra_gated_heading: DL3DV-10K terms acknowledgement required
extra_gated_prompt: >-
  The observation videos contain a DL3DV-derived 3DGS kitchen shell. Access is
  limited to users who have accepted the DL3DV-10K Terms of Use and will use the
  data only for non-commercial research or education.
extra_gated_fields:
  Full name: text
  Affiliation: text
  I have accepted the DL3DV-10K Terms of Use: checkbox
  I will use this material only for non-commercial research or education: checkbox

Canonical BiGym 3DGS Cutlery 32 dataset

This repository now has one canonical dataset identity only:

bigym-3dgs-light-floor-replay-plan-v2-20260802/dishwasher_unload_cutlery_long

It is the A800/CUDA LeRobot v3 collection selected by the owner, not the previously published Radeon/A800-parity split-file package.

Verified structure

• Task: DishwasherUnloadCutleryLong
• Episodes: 32 (episode_index 0–31)
• Successful saved episodes: 32/32, reward 1.0
• Frames: 21,018 at 20 fps
• State/action: finite float32 vectors of length 16
• Data: one merged Parquet file containing all 32 episodes
• Episode metadata: one merged Parquet file containing 32 rows
• Video: three merged H.264 streams, each 21,018 frames and 1,050.9 seconds
• Cameras: head 848×480, left wrist 640×480, right wrist 640×480
• Full local decode: passed for all three streams

One Parquet or MP4 chunk does not mean one episode. Episode boundaries are defined by episode_index and meta/episodes.

Canonical shell

Use only eustance/amd-bigym-3dgs-kitchen-shell, scene hash 90e70328...8947. The shell PLY and alignment hashes are locked by the dataset run-receipt.json.

Status boundary

Package structure, numerical integrity, replay rewards, strict renderer health, and video decoding pass. The source receipt still records visual_status: awaiting_visual_approval; this dataset is canonical by owner decision but is not relabeled as human visual-approved.

License and access

The observation videos contain an adapted DL3DV-10K scene. Non-commercial research and educational use only; recipients must accept the DL3DV-10K Terms of Use.