Optimization of humanoid’s motions under multiple constraints in vehicle ingress task

Abstract

This paper presents an approach on whole-body motion optimization for a humanoid robot to enter a ground vehicle. Motion capture system (mocap) was used to plan an initial suboptimal motion. Reinforcement learning was then implemented to optimize the trajectories with respect to kinematic and torque limits at the both body and the joint level. The cost functions in the body level calculated a robot’s static balancing ability, collisions and validity of the end-effector movement. Balancing and collision checks were computed from kinematic models of the robot and the vehicle model. Energy consumption such as torque limit obedience was checked at the joint level. Energy cost was approximated as joint torque, measured from a dynamic model. Various penalties such as joint angle and velocity limits were also computed in the joint level. Physical limits of each joint ensured both spatial and temporal smoothness of the generated trajectories. Finally, experimental evaluations of the presented approach were demonstrated through simulation and physical platforms in a real environment.