Probabilistic Inferences on Quadruped Robots: An Experimental Comparison

Abstract

Due to the reality gap, computer software cannot fully model the physical robot in its environment, with noise, ground friction, and energy consumption. Consequently, a limited number of researchers work on applying machine learning in real-world robots. In this paper, we use two intelligent black-box optimization algorithms, Bayesian Optimization (BO) and Covariance Matrix Adaptation Evolution Strategy (CMA-ES), to solve a quadruped robot gait's parametric search problem in 10 dimensions, and compare these two methods to find which one is more suitable for legged robots' controller parameters tuning. Our results show that both methods can find an optimal solution in 130 iterations. BO converges faster than CMA-ES within its constrained range, while CMA-ES finds the optimum in the continuous space. Compared with the specific controller parameters of two methods, we also find that for quadruped robot's oscillators, the angular amplitude is the most important parameter. Thus, it is very beneficial for the quick parametric search of legged robots’ controllers and avoids time-consuming manual tuning.