| --- |
| license: other |
| tags: |
| - robotics |
| - imitation-learning |
| - decentralized-control |
| - rgb-d |
| --- |
| |
| # RoboFactory Five-Task Decentralized Wrist RGB-D |
|
|
| Public, versioned research corpus for reproducible shared-policy experiments. |
|
|
| ## Contract |
|
|
| - `strict640x480-v2` contains 500 successful synchronized demonstrations: 100 each of LiftBarrier (2 robots), CameraAlignment (3), ThreeRobotsStackCube (3), LongPipelineDelivery (4), and TakePhoto (4). |
| - A policy stream contains only one `panda_hand` wrist RGB-D observation and that robot's qpos. |
| - RGB is native 640×480; depth is native metric depth stored in millimetres. Both encoders are audited to the aligned 30×40 patch grid (1200 tokens); historical 320×240 artifacts are deliberately excluded. |
| - No task ID, agent ID, peer/global/right-camera image, language, or communication is a policy input. |
| - HDF5 convention: action has `T` records, while RGB/depth/qpos have `T+1` records. The final observation is the post-action state. |
|
|
| ## Reproducible training protocol |
|
|
| All methods use the exact same corpus, held-out episode split, action chunk 100, |
| ACT posterior encoder 4 layers and decoder 7 layers. Mixed-task sampling is |
| hierarchical: equal probability for task, then demonstration, local arm, and |
| time index. The sampling label is never supplied to the policy. |
|
|
| The companion experiment uses frozen unseen 100-seed manifests per task and |
| publishes model configurations, normalization statistics, SHA-256 audits, and |
| results separately from this data repository. Each task is uploaded in its own |
| Hub commit only after `strict_audit.json` reports `PASS_STRICT640X480_V2`. |
|
|
