Deep Reinforcement Learning Based Collision Avoidance Algorithm for Differential Drive Robot

摘要

In this paper, collision avoidance problem is investigated for differential drive robot running in pedestrian environment, which requires for natural and safe interaction between robot and human. Based on deep reinforcement learning, a human-aware collision avoidance algorithm is proposed to find a smooth and collision-free path. A well designed reward function ensures the robot navigates without collision and obeys right-pass norm simultaneously. The slow convergence problem during training is addressed by pre-training the neural network using supervised learning. The simulation results show that the proposed algorithm can find a feasible and norm-obeyed path which achieves a natural human-robot interaction compared with traditional method.