Safe Robot Navigation with Reinforcement Learning using Dirichlet Distributions and Social Attention

Safe and socially aware navigation in human-populated environments remains a major challenge for autonomous mobile robots. This paper presents DIR-SAFE, a reinforcement learning-based local planner that integrates feasibility, safety, and social compliance for differential-drive robots. The method models the robot feasible velocity space using Dirichlet distributions and guarantees collision-free navigation via a lightweight action-space bounding algorithm informed by static obstacle maps. An actor-critic architecture with augmented state inputs enables efficient, real-time action inference without requiring online optimization or simulators. The policy is trained using Proximal Policy Optimization across diverse social interaction scenarios to promote robust generalization. Numerical results demonstrate that DIR-SAFE is an effective navigation algorithm, achieving high success rates and maintaining compliance with social-space constraints, even in dense and previously unseen environments, without requiring careful parameter tuning.