Adaptive deep reinforcement learning–based regenerative braking control for coordinated energy recovery and safety in electric vehicles

Regenerative braking control in electric vehicles is challenging due to the coupled influence of driver intent, following safety, and battery charge acceptance, making it difficult for conventional rule-based or static strategies to ensure both safety and energy recovery performance simultaneously. This paper proposes an adaptive constraint-aware deep reinforcement learning (ADRL) strategy for longitudinal motion control. The proposed framework unifies the battery charge acceptance limit and tire adhesion boundary into a physically feasible domain and embeds it into the policy learning process. An adaptive parameter mechanism is further introduced to estimate battery electrochemical–thermal states and to identify road adhesion conditions in real time, enabling online updates to the regenerative power upper bound and the slip safety threshold. Hardware-in-the-loop experiments under multiple cut-in scenarios show that ADRL increases the electric braking proportion by 17.3 and 5.1%, reduces hydraulic braking proportion by 22.0 and 6.4%, improves braking smoothness by 27.5 and 7.8%, and enhances cumulative energy recovery by 13.8 and 6.7% compared with rule-based and conventional DRL strategies. These results demonstrate the superior adaptability and efficiency of the proposed strategy under complex driving conditions.