Self Balancing Motorcycle Using Reinforcement Learning

This paper presents the application of Deep Deterministic Policy Gradient (DDPG) algorithm in Reinforcement Learning (RL) for motorbike self-balancing. The DDPG agent continuously interacts with the motorcycle environment to develop an optimum control policy without any knowledge about system dynamics. Agent takes states (observations) as input and output an action for which a delayed reward is received. The states of the motorcycle considered here are motorcycle position and velocity, while the action is the torque applied to the fly wheel motor. With the existing PD controller, the motorcycle’s orientation is adjusted by considering the deviation between its intended and actual position (Proportional component) and the pace of alteration of that Deviation (Derivative component). Here, the DDPG algorithm is used to become familiar with a policy that optimizes the reward signal based on the motorcycle’s stability. The proposed approach is evaluated through simulations and real-time experimentation, and the results show that the DDPG algorithm successfully learns a stable policy that can balance the motorcycle. Moreover, the proposed approach is robust to external disturbances, such as wind gusts. The results demonstrate the effectiveness of using RL algorithms in controlling complex dynamic systems, such as self-balancing motorcycles.