Model Hierarchy Predictive Control of Robotic Systems

This letter presents a new predictive control architecture for high-dimensional robotic systems. As opposed to conventional Model Predictive Control (MPC) approach locomotion that formulates hierarchical sequence of optimization problems, the proposed work single problem posed over hierarchy models, and is thus named Hierarchy (MHPC). MHPC formulated as multi-phase receding-horizon Trajectory Optimization (TO) problem, can be implemented using any general TO solver. benchmarked in simulation on quadruped, biped, quadrotor, demonstrating performance par or exceeding whole-body MPC while maintaining lower computational cost each case. A preliminary gap jumping experiment conducted MIT Mini Cheetah with policy generated offline, physical validity trajectories motivating online future work.