DexVision

Manipulation skills (reach, pick, place, push, press) for a simulated Shadow Hand in MuJoCo:

• Experts. Closed-loop scripted experts generate demonstrations.
• Learning. One small Diffusion Policy learns every skill.
• Physics. No orientation holds, no friction boosts, and success is measured on the physical hand and object.
• Evaluation. Eval-first: task spec, baselines and gates come before training, and every test-split run is logged.

The project is being rebuilt. See STATUS.md for where it stands and docs/PLAN.md for the checkpoints. The original pipeline is preserved at git tag legacy-v1; see docs/LEGACY.md for why it was retired.

Setup

conda env create -f environment.yml
conda activate dexvision
ruff check dexvision tests
pytest

On macOS, run viewer apps with mjpython. Automated tests need no GUI, GPU or camera.

Credits

The Shadow Hand model is from MuJoCo Menagerie (Apache-2.0; see assets/mujoco/menagerie/shadow_hand/LICENSE).