SNGNN2D-v2: A GNN-Based Model for the Generation of Human-Aware Cost Maps in Dynamic Environments

Abstract

Navigating dynamic, human-populated environments is a critical challenge for mobile robots, as they must balance effective pathfinding with minimizing social disruption. Cost maps can combine information from different nature and are more interpretable than final control signals. This paper addresses the generation of real-time cost maps in human-aware navigation (HAN) by introducing SNGNN2D-v2, a graph neural network designed and trained to capture social interactions and respond to dynamic elements in human-populated environments. SNGNN2D-v2 is evaluated through three types of experiments. The first involves deploying a real robot in a controlled indoor environment and assessing the disturbance caused by the robot when driven by the model. The second experiment tests the proposed model under more complex and unfavorable conditions using simulated environments. Both experiments include a comparison with other proposals using social and navigation metrics. The third experiment compares SNGNN2D-v2 with an end-to-end CNN-based method to evaluate how models generalize across changes in the appearance of the environment and its elements. The results from these experiments suggest that SNGNN2D-v2 is an effective model for human-aware cost map generation for dynamic environments. Its ability to capture dynamic information, generalize across scenarios with different appearances, and represent social interactions could contribute to the development of human-friendly robots.