Bayesian Optimization of a Quadruped Robot During 3-Dimensional Locomotion

摘要

Parametric search for gait controllers is a key challenge in quadruped locomotion. Several optimization methods can be adopted to find the optimal solution by regarding it as an optimization problem. Here we adopt Bayesian optimization (BO), a global optimization method that is suitable for unknown objective functions particularly when it is hard to evaluate, which is the common case of real robot experiments. We demonstrate this process on a quadruped robot capable of 3-dimensional locomotion, and our goal is to make it move forward as far as possible. While initially probing the parametric landscape, Random Search shows that in a 10-dimensional search space of over a million combinations, only 30% of them contribute to moving forward, merely 2% results in our robot walking longer than 2 m, and none of these parameters leads to more than 3 m distance. In face of such difficult landscape BO finds near-optimal parameters after 22 iterations, and walks a range of 3 m in over 40% of its iterations. Our findings illustrate that BO can efficiently search control parameters in a 3-dimensional locomotion case, and the development of controllers for legged robots, very often plagued with manual tuning of parameters, could profit from this.