Personalized Autonomous Braking for Electric Vehicles via Deep Reinforcement Learning and Learning From Demonstrations

Abstract This paper proposes an innovative reinforcement learning (RL)-based autonomous braking algorithm that can be personalizable for optimal one-pedal driving (OPD) of electric vehicles. To address the shortcomings of OPD–including its counterintuitive braking, which confuses drivers, causes fatigue and discomfort, and promotes a lack of conformity/trust–we propose a framework that integrates the state-of-the-art Twin Delayed Deep Deterministic Policy Gradient (TD3) RL agent with Learning from Human Demonstrations (LfD) via behavior cloning. An infusion term, λ, controls the influence of human demonstrations on policy shaping, allowing varying levels of personalization. For the RL agent, a comprehensive reward function is designed to balance precise braking, human comfort, and regenerative braking energy. Seven unique agents with different λ values are meticulously trained and evaluated against a baseline (λ=0) and a human-like (HL) algorithm in a full-braking scenario. The results show that incorporating a moderate value of human demonstration (λ=0.3) results in a more personalized and optimal control policy. Compared to the baseline (λ=0), the proposed agent achieves an improvement of 212% in precise braking and 0.3% in energy recovery, and a reduction of 24% in root-mean-square (RMS) jerk and 10% in human-like action dissimilarity. In comparison to the HL algorithm, the proposed agent shows an improvement of 0.4% in energy recovery and a reduction of 22% and 10% in RMS acceleration and jerk, respectively.