Reinforcement Learning-Based Autonomous Braking Control for One-Pedal Driving

Regenerative braking (RB) is vital for electric vehicles, enabling up to 9% energy recovery, especially with One-Pedal Driving (OPD). However, since OPD braking is counterintuitively controlled by the acceleration pedal, causing drivers to experience fatigue, discomfort, and confusion. This work addresses these issues by developing a reinforcement learning (RL)-based integrated autonomous braking algorithm using Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3) for OPD. The algorithm prioritizes precise braking, human comfort and maximizing regenerative braking, while decoupling braking from the acceleration pedal angle. Specifically, a comprehensive reward function is proposed that combines all the objectives required for optimal braking. Moreover, an ablation study is used to showcase the importance of each reward term. The trained agents are simulated in a full-braking scenario and compared against human driving data and a human-like (HL) algorithm– an aspect often overlooked in prior studies. The results show that our trained RL agents perform comparably to or better than human driving, with the TD3 agent showing superior performance over DDPG, and offering a 1.4% improvement in energy recovery and a 31.4% decrease in root-mean-square (RMS) deceleration, on average, compared to HL driving.