Social Navigation Among Workers and Robots in Dynamic Industrial Environments with Deep Reinforcement Learning

The growing deployment of autonomous mobile robots in factories and warehouses creates a need for navigation systems that ensure both physical safety and socially acceptable behavior when operating alongside human workers and other robots. Traditional rule-based and trajectory-predictive approaches can encode proxemic constraints but often lack adaptability in dynamic industrial environments with mixed traffic and structured layouts. This paper evaluates an attention-driven, learning-based social navigation framework that combines Deep Reinforcement Learning with Imitation Learning initialization, uses attention-based interaction encoding to prioritize relevant neighbors, and applies asymmetric social awareness that penalizes human discomfort while treating non-ego robots through collision avoidance only. Using consistent test-time metrics, we assess the trained policy across three industrially representative scenarios: open floor, fixed-layout environments with static obstacles, and a warehouse cross-intersection, and benchmark against a classical ORCA baseline under identical test conditions. The evaluation shows near-parity with ORCA in open and obstacle-rich layouts (open space: 0.97 vs 1.00 success; fixed layout: 0.82 vs 0.83 success), while improving robustness in the warehouse cross-intersection (0.26 vs 0.15 success, and reduced collision and timeout rates). The framework’s baseline training configuration achieved 89% success with 0.15% collision. Overall, the results highlight how structured industrial layouts amplify interaction difficulty, and they motivate scenario-based evaluation of socially aware navigation policies intended for deployment in smart mobility and logistics environments.