Multi-step recovery strategy for humanoid robots using model predictive control

Abstract

Humanoid robots in any environment are likely to experience collisions with obstacles or imbalance while attempting to navigate or perform a task. While humans are quite capable of balancing when encountering moderate collisions and imbalance, humanoid robots have yet to master the same skill. Push recovery is a strategy to maintain an upright posture in a robot while subjected to disturbances or imbalance. Performing push recovery on a humanoid robot consists of planning the actuation of the joints, ground reaction forces, and footsteps. Controlling a robot's trajectory has many challenges such as incorporating nonlinear dynamic equations, achieving resilience to disturbances, and maintaining computational efficiency. This paper introduces a control framework that determines the necessary actuated joint forces and footsteps in order to balance. The proposed control framework is distinct from existing Capture Point/Capture Region solutions in that the controller simultaneously plans multiple future footsteps and all actuated joint forces. Model Predictive Control provides the necessary planning and optimization to anticipate the robot's future trajectory and accommodate imbalance. Combining the effort of all actuators and multiple footsteps into one recovery strategy has yet to be implemented in push recovery of humanoid robots. This control framework is demonstrated with a bipedal walking robot model and the performance is validated with simulation results.