The framework uses CF-Gen to synthesize contact-flow segments and CF-Track to execute them with a low-level policy. Closed-loop monitoring runs at 50 Hz, while a live MuJoCo WASM and ONNX viewer exposes the policy in the browser. A VLM integration translates natural-language goals into multi-step manipulation sequences and the project includes recovery behaviors for long-horizon execution.
OmniContact is useful for humanoid-robot research, sim-to-sim validation, manipulation planning, and benchmarking compositional skills. Its public dataset includes 74,641 MoCap clips, 1,274 valid paired sequences, 22.29 hours of motion, and 7.22 million object frames, providing artifacts for studying contact-rich behavior and generalization.
