armanual — bimanual SO-101 dinner-table demonstrations
Demonstrations for Bimanual VLA Manipulation with Multi-Modal Reasoning (Intel Physical AI Online Challenge). Two simulated SO-101 arms set a dinner table in MuJoCo: placing plates, cups and cutlery, opening a drawer to reach the cutlery, handing objects between arms, and pouring.
Code: https://github.com/pythonsniffer/armanual
What is in it
| Episodes | 114 |
| Frames | 22,206 (20 Hz) |
| Distinct instructions | 30 |
| Robot | dual SO-101 (so101_bimanual) |
| Format | LeRobotDataset v3.0, video-encoded |
Features
| Column | Shape | Meaning |
|---|---|---|
observation.images.top |
224×224×3 | overhead camera |
observation.images.left_wrist |
224×224×3 | left arm wrist camera |
observation.images.right_wrist |
224×224×3 | right arm wrist camera |
observation.state |
12 | 5 joints + gripper opening, per arm (left first) |
action |
12 | commanded joint targets, same layout |
task |
text | the instruction sentence that produced the episode |
Actions are bimanual. Both arms are recorded at every step, not just the moving one — hand-offs and hold-while-pouring depend on what the other arm is doing, and recording a single arm makes those behaviours structurally unlearnable.
Skills
| Skill | Episodes | Example instruction |
|---|---|---|
place_object |
87 | "put the blue cup to the right of the plate" |
handoff |
14 | "pass the cup to the other arm" |
open_drawer |
13 | "open the cutlery drawer" |
pour |
0 | (the expert's pour succeeded too rarely to contribute demonstrations) |
Each skill has several phrasings so a policy keys on meaning rather than a memorized string.
How it was generated
Episodes are produced by running an analytical controller — camera-based perception, referent grounding, torque- and collision-screened IK, scripted manipulation primitives — on generated instructions, recording every control tick. The instruction that produced an episode is its label.
Only successful episodes are included. Success is checked against the final state of the table (was the object actually placed where the sentence said?), not against the controller's own opinion. The expert succeeds about 55% of the time; the failures are discarded.
Randomization during collection: object placement, size and mass. Lighting, friction, colour and clutter were held fixed, because enabling them drops the expert's success rate to ~25% and would have shrunk the dataset threefold. Policies trained on this data are evaluated under the full randomization, which is where generalization should be measured.
Known limitations
- No pour demonstrations. The expert's bottle grasp is its weakest skill; every pour attempt in the collection run failed its success check and was discarded.
- Visual domain is narrower than the evaluation domain (see above).
- The demonstrations inherit the expert's failure modes; a policy trained on them should not be expected to exceed it on skills where the expert is weak.
Reproducing it
git clone https://github.com/pythonsniffer/armanual && cd armanual
pip install -e ".[dev]" && pip install "lerobot[smolvla]"
python scripts/collect_dataset.py --episodes-per-skill 60 --workers 6
Licence
Apache-2.0. The SO-101 model is from mujoco_menagerie (Apache-2.0).
- Downloads last month
- 190