MAPPO-Based Optimal Reciprocal Collision Avoidance for Autonomous Mobile Robots in Crowds

This paper proposes an improved Optimal Reciprocal Collision Avoidance (ORCA) algorithm for robot navigation in crowds using a deep reinforcement learning algorithm, Multi-agent Proximal Policy Optimization (MAPPO). The original ORCA algorithm allows the robot to compute escape velocities to escape from collisions with humans and then gets a set of collision-free velocities based on the reciprocal assumption. But humans usually don't follow the assumption, and thus ORCA performs badly. In this paper, we propose a MAPPO-based method to explore the optimal escape velocity for the robot to each human and use an estimation module to decide the effect of each escape velocity on the robot's new velocity. The proposed method is evaluated in simulated crowds of humans, and the original ORCA algorithm with different configurations are used as baselines. Simulation results show the proposed method achieves a significantly higher navigating success rate with almost the same time consumption compared with the original ORCA algorithm in robot navigation among crowds.