Group and Socially Aware Multi-Agent Reinforcement Learning

Abstract

Many researches in the field of robot navigation show the effectiveness of Deep Reinforcement Learning and Reward Function Modeling for Crowd Navigation and Multi-Agent Reinforcement Learning. The notion of groups has not yet been studied in the context of Reinforcement Learning. A robot using the current approaches is likely to walk in-between a group of people, while a robot moving alongside with a group of people is unlikely to make an extra effort to avoid group splitting when avoiding other people. We learn the behavior of multiple-robots to be group-aware to avoid breaking of the groups, while also being-socially aware to leave comforting personal space from the other people. The work uses Imitation Learning on a dataset produced by using the Social Potential Field algorithm to kick start the learning of the Reinforcement Learning policy. The learning is facilitated by the reward function that is specifically modelled to learn the desired behaviours. The proposed work is compared against the Artificial Potential Field Algorithm, Social Potential Field Algorithm, Optimal Reciprocal Collision Avoidance and Reinforcement Learning baselines and found to be the best among all these approaches.