Deep Reinforcement Learning for Group-Aware Robot Navigation in Crowds

Balancing safety, efficiency and social norms when robots navigate in crowded situations is a major challenge, as human intentions are variable and unpredictable. Previous research has addressed the problem of navigating in dense crowds by modelling the crowd and using a self-attention mechanism to assign different weights to each individual. However, in reality, crowds do not only consist of individuals, but more often appear as groups, so avoiding groups is also part of social etiquette. In particular, humans are aware of other groups and actively circumvent them when they keep a certain distance from others. Based on this, we proposes a deep reinforcement learning navigation method that incorporates group recognition. The proposed method distinguishes groups from individuals by behavioural similarity and designs a group space-based reward function in a reinforcement learning framework to encourage the robot to avoid group intrusion. According to the simulation experiments, the suggested method can produce navigation paths that are more group-friendly.