DRL-Based Adaptive Planner Switching for Mobile Robots: CPN Modeling, Verification and Optimization

Mobile robots operating in complex environments must switch planners as road conditions change, yet concurrent triggers make the switching logic difficult to validate through code inspection and empirical tuning alone. We use Coloured Petri Nets (CPN) to analyze a four-mode scheme, expose unbounded transition-buffer accumulation and rollback oscillation, and derive a Guide–Safe–Emergency architecture with explicit responsibilities, unified arbitration, and single-channel execution. Exhaustive exploration of its finite interface instance covers 3,486 reachable markings and 13,097 arcs; all five specified correctness properties are satisfied. Within this architecture, we adopt a safety-constrained reinforcement learning approach: the DQN is authorized only to optimize Guide–Safe switching based on recent environmental complexity and switching history, whereas a non-learning Supervisor and an independent Safety Guard enforce mode mutual exclusion, priority arbitration, Emergency intervention, and conservative recovery. The DQN cannot alter or bypass these constraints. In simulation, the constrained policy improves average cumulative reward by 10.1% over a fixed-threshold baseline. Deployment on an Ackermann-steering robot demonstrates multi-goal planner switching and two closed-loop Emergency stop–recovery cycles, confirming hardware executability of the CPN-derived constrained-AI architecture.