Crowd-Aware Socially Compliant Robot Navigation via Deep Reinforcement Learning

Navigating in crowd environments is challenging for mobile robots because not only the safety but also the comfort of surrounding pedestrians must be considered. In this work, a deep reinforcement learning framework is introduced for safe and socially compliant robot navigation. We propose a value network for robot decision-making that leverages spatial-temporal reasoning to comprehend crowd interactions. Based on the real-time speed of pedestrians, the hazardous areas that the robot needs to avoid are designed, and a reward function is formulated to guarantee the safety and comfort of pedestrians. Extensive simulation experiments validate that the developed framework outperforms the state-of-the-art methods in terms of success rate (up to 34% increase) and discomfort frequency (up to 54.72% decrease). In addition, real-world experiments illustrate that our approach can predict pedestrian dynamics and navigate the robot safely and reliably in crowds.