Efficient reinforcement learning for humanoid whole-body control

摘要

Whole-body control of humanoid robots permits the execution of multiple simultaneous tasks but combining tasks can often result in unexpected overall behaviors. These discrepancies arise from a variety of internal and external factors and modeling them explicitly would be impractical. Reinforcement learning can be used to eliminate the effects of the deleterious factors through trial and error but generally requires many trials to converge on a solution. In humanoid robotics such improvidence can be costly. In this paper we show how the efficiency of the learning can be improved through use of Bayesian optimization. This is accomplished by intelligently exploring a model of the latent cost function derived from the quality of the task executions. We demonstrate the efficacy of the technique through two different simulated scenarios where various factors impede the robot from accomplishing its objectives.