Robust Robot Control Using Multiple Model-Based Policy Optimization and Fast Value Function-Based Planning.

摘要

This report describes the research findings of Team Steel, the group led by PI Christopher G. Atkeson in the DARPA Virtual Robotics Challenge (VRC). They developed human-like walking and robust walking on rough terrain, and automated driving in the VRC context. They developed a rough terrain footstep planner, a decoupled approach to state estimation, and an optimization based real-time walking controller for a full size 3D humanoid robot. They showed that optimal stepping trajectories and trajectory cost for a walking biped robot on rough terrain can be encoded as simple quadratic functions of initial state and footstep sequence.