Motion Planning with Slip Prediction Based on Reinforcement Learning

This paper first reformulates the motion planning problem as a continuous decision-making task, making it suitable for reinforcement learning. It then develops a network to predict traffic costs, integrating this network with a global planning algorithm as a heuristic function. Simultaneously, an action network is designed, which takes local objectives and environmental information as input. By leveraging the experience gained during motion, the proposed algorithm enhances both the accuracy and generalization of slip prediction. It generates actions that mitigate the effects of slippage, ensuring the actual path closely aligns with the planned trajectory, thereby minimizing rover deviations. An environmental model is constructed for training and validation to assess the effectiveness of the proposed approach.