Close-range docking control for reconfigurable ground vehicles: Model-guided reinforcement learning with robust predictive safety filter

Reconfigurable ground vehicles (RGVs) equipped with all-wheel independent steering (AWIS) provide enhanced mission adaptability but present challenges for achieving high-precision autonomous docking, particularly during the close-range capture stage (CCS). This paper presents a novel control strategy for close-range docking control based on the twin delayed deep deterministic policy gradient (TD3) and robust predictive safety filter (RPSF). The key artificial intelligence (AI) contribution lies in a model-guided reinforcement learning (MGRL) training framework that leverages prior optimal control solutions derived from the generalized v-β-r RGV dynamic model to accelerate the TD3 learning process. This strategy employs theoretical steering radius angle θ_R and sideslip β_R angle as generalized control inputs, dynamically adjusting the instantaneous center of rotation (ICR) to fully leverage AWIS’s capabilities for precise position and orientation control. The engineering application focuses on the integration of a RPSF with the trained reinforcement learning (RL) agent for enhanced CCS docking control of the RGV. The RPSF incorporates feedback compensation to mitigate model-plant mismatch and employs online optimization to ensure strict compliance with state and control constraints throughout the docking process. Simulation results validate the proposed method, demonstrating millimeter-level docking accuracy in CCS while effectively utilizing AWIS agility and guaranteeing safety constraint satisfaction.