PPO-GAT-Follow: Graph-Attention Reinforcement Learning for Robust Robot Person Following in Dense Crowds

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Robot person following (RPF) in dense crowds is challenging due to navigation complexities and social constraints.
• Maintaining target visibility and appropriate relative positioning requires sophisticated algorithms.

Purpose of the Study:

• To propose an interaction-aware reinforcement learning framework, PPO-GAT-Follow, for dense-crowd RPF.
• To address geometric visibility loss while utilizing target-relative pose estimates.
• To ensure robust performance under various crowd densities and dynamic conditions.

Main Methods:

• Utilizing a graph attention encoder to model interactions between the follower, target, and surrounding pedestrians.
• Implementing a task-oriented reward mechanism encompassing target maintenance, visibility, collision avoidance, and social compliance.
• Conducting experiments in simulation (IR-SIM) and validation in Gazebo.

Main Results:

• PPO-GAT-Follow achieved 98.8% task success and 1.1% collision rate in a fixed-route setting, outperforming MPC by 10.9%.
• In random-route settings, it reached 83.1% task success and an SPL of 0.815, surpassing MPC by 18.3%.
• Zero-shot evaluations and stress tests confirmed the framework's effectiveness and reliability across diverse scenarios.

Conclusions:

• PPO-GAT-Follow demonstrates superior performance in dense-crowd RPF compared to existing methods.
• The framework's interaction-aware approach and comprehensive reward mechanism are key to its success.
• System integration feasibility is validated, though visual identification and occlusion recovery require further research.