control/
Dual-mode locomotion (wheel + leg), dock-approach controller, manipulation policy, and terrain classification for Free Humanoid Centaur.
This directory will hold:
- Wheel-mode skid-steer kinematics + per-wheel-foot PID. Four-wheel skid-steer (state: x, y, heading, linear and angular velocity; control: per-wheel torque). Cycle 200–500 Hz on the control compute domain.
- Wheel-mode MPC. Horizon ~1 s, cycle ~50 Hz. Constraints: linear velocity ≤ 2.0 m/s, angular velocity ≤ 1.5 rad/s. Hands off to the safety supervisor on slope-tilt or edge-proximity violations.
- Leg-mode model-based MPC. Horizon ~1 s, cycle ~500 Hz. Standard quadruped MPC stack —
mit-cheetah-2, anymal, mini-cheetah lineage.
- Leg-mode RL perceptive policy. Sim2real RL policy at ~50 Hz for foothold selection and terrain-aware locomotion. Shielded by
hwangbo-anymal-sim2real.
- Mode-transition controller. State machine implementing the hard-mode-switch sequence: bring-to-zero, latch-current-pose, verify-IMU-tolerance, reconfigure-actuator-roles, release-brakes. Coordinates with the safety supervisor’s mode-transition gate invariant.
- Terrain classification. Lidar ground-plane processing on the perception domain; outputs surface labels (paved, gravel, mud, vegetation, water, stair, edge) used by the mode-selection policy.
- Dock approach controller. Fine-positioning relative to the dock-A optical fiducial (head stereo camera + dock-burst optical handshake from shoal ARCHITECTURE §4). Targets ±20 mm positional accuracy at the cartridge port. Same accuracy spec as the wheeled sibling’s dock-B approach.
- Manipulation policy. Inherited from the bipedal sibling — imitation-learning baseline (ACT or diffusion-policy variant) for cartridge-swap and waste-hopper retrieval.
- Whole-body MPC for upper-body manipulation while in
manipulation_static mode. Inherits the bipedal sibling’s whole-body MPC reduced to upper-body-only.
Cross-references
Status
v0.1 scaffold. The dual-mode controller — wheel-mode skid-steer MPC + leg-mode model-based MPC + leg-mode RL perceptive policy + mode-transition state machine — is the Phase 1 deliverable (closed-loop in MuJoCo).