A reinforcement learning-enhanced collision avoidance strategy integrating motion planning and control

Existing automotive collision avoidance (CA) systems typically adopt a multi-layer architecture, separating motion planning and control due to computational complexity and differing optimization horizons, leading to potential conflicts and compromised safety in complex collision avoidance scenarios. To this end, this paper proposes a novel reinforcement learning-enhanced CA strategy that integrates motion planning and control processes. A cascaded nonlinear model predictive control (CNMPC) approach is proposed, employing high-fidelity nonlinear dynamics for immediate control actions and simplified linear dynamics for long-term trajectory planning. Additionally, a horizon-decreasing risk-weighting (HDRW) strategy is introduced to prioritize imminent collision risks, addressing the limitations of traditional model predictive control (MPC) in multi-obstacle situations. Furthermore, reinforcement learning (RL)-based adaptive control agents are developed to dynamically adjust the coordination strategy between the longitudinal and lateral inputs based on real-time road adhesion, enhancing the system’s environmental adaptability and robustness. Simulations and hardware-in-the-loop (HIL) experiments under multi-obstacle CA scenarios with low and varying road adhesion demonstrate the proposed strategy’s feasibility and superior performance, ensuring safe and stable collision avoidance under complex road conditions.