Deep Reinforcement Learning-Based Control of Bicycle Robots on Rough Terrain

Abstract

The bicycle robot is a kind of unmanned mobile robot with great potentials. However, the control of such robots on rough terrain under model uncertainties and disturbances is challenging due to the lateral instability and underactuated char-acterisitcs. In this paper, a controller based on deep reinforcement learning and Stanley algorithm is designed to achieve the path tracking and balancing control for a bicycle robot on rough terrain. First, the outputs of the controller are compensated by residual reinforcement learning to reduce the gap between the training environment and the test environment. Then, in balancing control, the piecewise curriculum learning is designed to accelerate the training process. In path tracking control, Stanley algorithm is used. The proposed algorithm is tested in Gazebo simulation environment. The results showed that the proposed algorithm achieved good performance.