Using Dynamic Rewards to Learn a Fully Holonomic Bipedal Walk

摘要

This paper presents the design and learning architecture for a fully holonomic omnidirectional walk used by the UT Austin Villa humanoid robot soccer agent acting in the RoboCup 3D simulation environment. By“fully holonomic” we mean the walk allows for movement in all directions with equal velocity. The walk is based on a double linear inverted pendulum model and was originally designed for the actual physical Nao robot. Parameters for the walk are optimized for maximum speed and stability while at the same time a novel approach of reweighting rewards for walking speeds in the cardinal directions of forwards, backwards, and sideways is utilized to promote equal walking velocities in all directions. A variant of this walk which uses the same walk engine, but is not fully holonomic as it employs three different sets of learned walk parameters biased toward maximizing forward walking speed, was the crucial component in the UT Austin Villa team winning the 2011 RoboCup 3D simulation competition. Detailed experiments reveal that adaptively changing the weights of rewards over time is an effective method for learning a fully holonomic walk. Additional data shows that a team of agents using this learned fully holonomic walk is able to beat other teams, including that of the 2011 RoboCup 3D simulation champion UT Austin Villa team, that utilize non-fully holonomic walks.