Leveraging Interaction Uncertainty for Enhanced Navigation Performance in Dynamic Crowds

Tao Zou, Haonan Yang, Ping Zhong

Abstract: In dynamic human crowds, robot navigation faces significant challenges due to uncertainty, which can negatively impact both navigation success and human safety. Although some approaches have attempted to address this issue, they often focus on modeling uncertainty at the individual pedestrian level, over-looking the complex interactions between pedestrians and robots. To tackle this gap, this paper introduces an adaptive, data-driven uncertainty interaction perception network that accounts for uncertainties in pedestrians, robots, and their mutual behaviors. By assigning adaptive weights to different attention heads, the network dynamically highlights key interaction features, enabling the robot to manage varying levels of uncertainty effectively. This dynamic adjustment enhances the robot’s ability to navigate unpredictable environments while maintaining a balance between safety and efficiency. Additionally, to better reflect the impact of uncertainty, our reward function incorporates the Mahalanobis distance, replacing the commonly used Euclidean distance. This provides a more accurate assessment of potential collision risks, enhancing both safety and social awareness. Through extensive experiments, we show that our approach outperforms state-of-the-art baselines in terms of safety, adherence to social norms, navigation efficiency, and robustness in extreme environments.

External IDs:dblp:conf/ijcnn/ZouYZ25