free-humanoid-platform

control/

Whole-body MPC reference, RL policy weights, manipulation policy.

Status (v0.0)

Stub.

Architectural posture

Layered control:

1. Low-level (200–500 Hz): whole-body MPC. Textbook formulation (no patentable claims). Implemented in Rust per OpenLoco’s gait_harness posture or in C++ if leveraging an existing MPC library (Crocoddyl, Aligator). Outputs per-actuator impedance commands.

2. Mid-level (50–100 Hz): RL policy for locomotion mode and gait. Trained in MuJoCo via OpenLoco’s gait_harness (the same harness that runs the existing ANYmal-WL and Mini-Whegs gaits). Outputs MPC reference trajectories.

3. High-level (10 Hz): manipulation policy. Imitation-learning baseline (ACT or diffusion-policy variant). Optional VLA upgrade. Outputs target end-effector trajectories.

Prior-art shielding

See ../prior-art/INDEX.md §4:

• MPC: textbook (no specific corpus citation needed; mit-cheetah family establishes the open-MPC tradition for legged platforms).
• RL: mit-cheetah-2, cassie-osu, mini-cheetah, digit-meta, berkeley-humanoid.
• IL: act-aloha, mobile-aloha, diffusion-policy.
• VLA: openai-rt-2, open-x-embodiment, physical-intelligence-pi-zero, skild-foundation-model, covariant-rfm.

Planned content

• mpc/ — whole-body MPC reference
• rl/ — RL policy training scripts and weight checkpoints
• il/ — IL policy training scripts and weight checkpoints
• vla/ — optional VLA wrapper (subprocess to a foundation-model API or local model)
• eval/ — evaluation harnesses

License

Apache-2.0 (per ../LICENSE-SOFTWARE) for code. CC0-1.0 (per ../LICENSE-DATA) for trained policy weights treated as data.

This site is open source. Improve this page.