AI Robots Navigate Crowds Socially with Proxemics-Based Rewards

Developing robots that can navigate effectively and acceptably in human-dense environments is a critical challenge for real-world AI applications. Current deep reinforcement learning (DRL) methods for navigation often prioritize task completion, sometimes at the expense of social compliance. This paper proposes a new reward formulation designed to integrate social norms into DRL-based robot navigation. The core of the approach involves modeling each human's personal space based on Hall's proxemics theory, using a radial Gaussian-mixture field. This allows the robot to compute a local cost within its field of view, providing a dense and interpretable social learning signal. By incorporating this proxemics-based reward into existing DRL navigation methods, the researchers demonstrated in simulations that robots could significantly improve their social metrics—such as maintaining appropriate distances and avoiding uncomfortable proximity—while still achieving competitive navigation performance across various crowd scenarios and densities.