A Review of Self-balancing Robot Reinforcement Learning Algorithms

Abstract

We analyse reinforcement learning algorithms for self balancing robot problem. This is the inverted pendulum principle of balancing robots. Various algorithms and their training methods are briefly described and a virtual robot is created in the simulation environment. The simulation-generated robot seeks to maintain the balance using a variety of incentive training methods that use non-model-based algorithms. The goal is for the robot to learn the balancing strategies itself and successfully maintain its balance in a controlled position. We discuss how different algorithms learn to balance the robot, how the results depend on the learning strategy and the number of steps. We conclude that different algorithms result in different performance and different strategies of keeping the robot balanced. The results also depend on the model training policy. Some of the balancing methods can be difficult to implement in real world.