Add files using upload-large-folder tool
Browse files- README.md +32 -0
- corpus_manifest.json +27 -0
README.md
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
private: true
|
| 4 |
+
tags:
|
| 5 |
+
- robotics
|
| 6 |
+
- imitation-learning
|
| 7 |
+
- decentralized-control
|
| 8 |
+
- rgb-d
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# RoboFactory Five-Task Decentralized Wrist RGB-D
|
| 12 |
+
|
| 13 |
+
Private research corpus for reproducible shared-policy experiments.
|
| 14 |
+
|
| 15 |
+
## Contract
|
| 16 |
+
|
| 17 |
+
- 500 successful synchronized demonstrations: 100 each of LiftBarrier (2 robots), CameraAlignment (3), ThreeRobotsStackCube (3), LongPipelineDelivery (4), and TakePhoto (4).
|
| 18 |
+
- A policy stream contains only one `panda_hand` wrist RGB-D observation and that robot's qpos.
|
| 19 |
+
- RGB is 640×480; depth is native metric depth stored in millimetres.
|
| 20 |
+
- No task ID, agent ID, peer/global/right-camera image, language, or communication is a policy input.
|
| 21 |
+
- HDF5 convention: action has `T` records, while RGB/depth/qpos have `T+1` records. The final observation is the post-action state.
|
| 22 |
+
|
| 23 |
+
## Reproducible training protocol
|
| 24 |
+
|
| 25 |
+
All methods use the exact same corpus, held-out episode split, action chunk 100,
|
| 26 |
+
ACT posterior encoder 4 layers and decoder 7 layers. Mixed-task sampling is
|
| 27 |
+
hierarchical: equal probability for task, then demonstration, local arm, and
|
| 28 |
+
time index. The sampling label is never supplied to the policy.
|
| 29 |
+
|
| 30 |
+
The companion experiment uses frozen unseen 100-seed manifests per task and
|
| 31 |
+
publishes model configurations, normalization statistics, audits, and results
|
| 32 |
+
separately from this private data repository.
|
corpus_manifest.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"version": "five_task_rgbd_v1",
|
| 3 |
+
"tasks": {
|
| 4 |
+
"lift_barrier": {
|
| 5 |
+
"episodes": 100,
|
| 6 |
+
"source": "/workspace/RoboFactory/runs/wrist_rgbd640/lift_barrier"
|
| 7 |
+
},
|
| 8 |
+
"camera_alignment": {
|
| 9 |
+
"episodes": 100,
|
| 10 |
+
"source": "/workspace/RoboFactory/runs/wrist_rgbd640/camera_alignment"
|
| 11 |
+
},
|
| 12 |
+
"three_robots_stack_cube": {
|
| 13 |
+
"episodes": 100,
|
| 14 |
+
"source": "/workspace/RoboFactory/runs/wrist_rgbd640/three_robots_stack_cube"
|
| 15 |
+
},
|
| 16 |
+
"long_pipeline_delivery": {
|
| 17 |
+
"episodes": 100,
|
| 18 |
+
"source": "/workspace/RoboFactory/runs/wrist_rgbd640/long_pipeline_delivery_atomic_canary"
|
| 19 |
+
},
|
| 20 |
+
"take_photo": {
|
| 21 |
+
"episodes": 100,
|
| 22 |
+
"source": "/workspace/RoboFactory/runs/wrist_rgbd640/take_photo"
|
| 23 |
+
}
|
| 24 |
+
},
|
| 25 |
+
"policy_input": "single local panda_hand wrist RGB-D plus local qpos",
|
| 26 |
+
"depth_unit": "millimeters"
|
| 27 |
+
}
|