Robot-Crowd Navigation with Socially-Aware Reinforcement Learning Over Graphs

Abstract

Robots typically perform navigation task in a crowd environment, where the navigation task requires robots to reach a target point safely and efficiently, and to have the least impact on crowd trajectories. To this end, we propose a graph-based socially aware reinforcement learning navigation algorithm, in which the robot-crowd interactions are modeled as a directed spatio-temporal graph. We utilize graph convolutional networks, attention mechanism and long short term memory networks to encode robot-crowd interaction features, which are subsequently leveraged for state value estimation and robot action selection. Our method is demonstrated to have high success rate and short navigation time in various environments and outperform existing methods in terms of security and efficiency.