| --- |
| license: other |
| tags: |
| - robotics |
| - imitation-learning |
| - decentralized-control |
| - multi-robot |
| - rgb-d |
| --- |
| |
| # RoboFactory Five-Task Decentralized Wrist RGB-D |
|
|
| Public research corpus for reproducible Stereo-CoRE shared-policy experiments. |
|
|
| ## Contract |
|
|
| - 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 640x480; depth is native metric depth stored in millimetres. |
| - 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 same corpus, held-out episode split, action chunk 100, ACT posterior encoder |
| with 4 layers, and policy decoder with 7 layers. The main Stereo-CoRE run uses hierarchical |
| weighted item sampling: equal probability for task, then demonstration, local arm, and time index. |
| The sampling label is never supplied to the policy. |
|
|
| ## Companion releases |
|
|
| - Code and protocols: https://github.com/YananZHOU5555/Stereo-CoRE |
| - Weights, configurations and raw results: https://huggingface.co/B111ue/Stereo-CoRE |
|
|
| The repository-level `corpus_manifest.json` and per-task audit/manifest files describe the public |
| payload. Users remain responsible for complying with RoboFactory and simulator asset licenses. |
|
|
