A Safety-Constrained Reinforcement Learning Model for Autonomous Navigation in Cluttered Industrial Spaces

Abstract

Navigation in dynamic industrial environments for robotic systems is particularly problematic due to the classification of obstacles, safety policies, spatial constraints, and the presence of constantly changing dynamic entities. In this work, we propose a novel solution to constraint path planning using safety-grained reinforcement learning where lower bounds on safety regarding control actions within the plant are set a priori. Collision avoidance and proximity bounds during reward shaping policy optimization, avoidance, and proximity violation unlock exceeding defined safety constraints. The framework is tested with industrial layouts of varying complexity and density of obstacles, which have been simulated and field tested. Results achieved within RL baseline models surpass the thresholds set for the defined safe bounding box for pathing and collision avoidance, demonstrating increased efficiency. Alarms encapsulated within the model maintain consistent robustness under arbitrary and unanticipated tracking scenarios where randomly placed obstacles may move during unobstructed times across unpredicted locations. Adaptive learning policies bolster traditional frameworks at the guarded learning core time constraints, allowing for responsive violation of safety regions in a real-time response to changing warehouse and industrial environments. Further developments expand future objectives by combining multi-agent coordinated targeting situational awareness with semantic mapping.