Deep Reinforcement Learning Based Crowd Navigation via Feature Aggregation from Graph Convolutional Networks

摘要

In this paper, we use the graph convolutional network (GCN) for feature aggregation. Our approach, termed as GCN-RL, can directly deploy on a holonomic mobile robot without any tuning. We first use GCN to extract the hidden features among the robot and humans. These extracted features that represent the spatial relationships and agents-agents interactions are then fed into the actor-critic learning framework. Finally, the deep RL network is optimized based on the aggregated features from GCN and the actor-critic framework. The GCN-RL enables a safer and more efficient navigation policy than the other RL navigation methods. The experiment results show that the proposed learning approach significantly outperforms ORCA and other RL navigation methods.