Perception-Aware Multi-Terrain Humanoid Walking via Nonlinear MPC
·
1 min read
- Developed an nonlinear MPC planner for perception-aware multi-terrain humanoid walking based on OCS2, using a centroidal dynamics model that incorporates joint positions and velocities to account for limb dynamics and foothold region constraints, and replanning at 100 Hz.
- Enhanced the integration of perception and control by explicitly considering foothold region constraints in the planning process.
- To address the computational complexity of nonlinear MPC, developed a task-space QP-based whole-body control strategy that enabled 1 kHz torque-based whole-body control.
- Validated the approach in RaiSim simulations and on the Wukong-IV humanoid robot, increasing maximum forward speed from 1.0 m/s to 1.7 m/s and enabling traversal of 22° slopes and 25 cm steps in simulation.
Authors
Kunzhao Ren
(he/him)
PhD Student
Hi, I am Kunzhao Ren, a first-year PhD student in Mechanical Engineering at the University of Wisconsin–Madison, with interests in legged robotics.