Balance Control of an Inverted Pendulum on a Quadruped Robot by Reinforcement Learning

Abstract

When a quadruped robot crosses some uneven terrain, its stability is a very important consideration. In practical applications, it is often difficult for researchers to obtain accurate models of quadruped robot systems and need quadruped robots to complete high-stability tasks. Balancing an inverted pendulum on a quadruped robot is a good way to test their balancing ability, which is of great research and practical value. According to the nonlinear, uncertain, and strong coupling characteristics of the quadruped robot inverted pendulum system, this paper proposes a quadruped robot balancing inverted pendulum algorithm based on reinforcement learning, the Q-learning algorithm. The advantage of this algorithm is that it can learn effective balancing policy directly from experience, and it does not depend on the accurate model of the quadruped robot. It has the characteristics of high efficiency and flexibility. Through a lot of training, it can break through the performance limit brought by model error to traditional methods. In this paper, a comparative experiment of a set of balanced inverted pendulums of quadruped robots with different sizes is completed in the V-REP simulation software. The experimental results show that the algorithm can effectively improve the balance ability of quadruped robots, and it also shows that the algorithm has good adaptability.