Abstract : Efficient integration of Deep Reinforcement Learning (DRL) into robotic systems faces challenges due to the computational complexity and vast parameter requirements of traditional models. This paper introduces a novel framework combining DRL with heuristic algorithms to simplify command generation for Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) in manufacturing tasks. By decomposing system states and actions into distinct categories—time scheduling, task assignment, and trajectory planning—the proposed approach employs lightweight Deep Q-Networks and the Dijkstra algorithm for optimization. This design minimizes computational overhead, accelerates convergence, and reduces memory usage while ensuring effective task execution. Numerical evaluations highlight the efficiency gains of the simplified DRL model over conventional approaches, showcasing a significant reduction in parameters, training time, and inference latency. The findings demonstrate the potential for this modular optimization strategy to enhance the performance of autonomous systems across diverse domains.

Keyword : Deep Reinforcement Learning, heuristic algorithms, low complexity, UAV, UGV.