Semi-Supervised Deep Reinforcement Learning in Decentralized Multi-Agent Collision Avoidance and Path Planning in a Complex Environment

Abstract

The problem of path planning and collision avoidance in complex and natural environments is one of the basic requirements of the robotic world, enabling robots to enter social environments. This paper aims to provide a decentralized path planning and collision avoidance method in multi-agent environments. In this method, each agent is a decision-making unit that decides independently from other agents and based on what is in its field of view. In the present paper, classical methods have been used to generate data for training purposes. Models were trained offline by imitating classical methods then semi-supervised methods were used for feature extraction. The results obtained from this method were compared with the Optimal Reciprocal Collision Avoidance (ORCA) method in three environments with different densities and three different indices. The proposed method performed relatively optimally and successfully increased the interaction index while decreasing the computation time. On the other hand, due to the scalable potential of this method, the number of agents could be increased without affecting the computation time.