Safe, efficient, and energy-saving speed control for autonomous driving based on reinforcement learning
This paper proposes a deep reinforcement learning framework for automated eco-driving in car-following scenarios. The framework uses Twin Delayed Deep Deterministic Policy Gradient (TD3) to optimize safety, traffic efficiency, and electric-powertrain operating efficiency. A physics-informed electric vehicle powertrain model is used to evaluate motor efficiency, and a multi-objective reward combines Time-to-Collision (TTC), time headway, and powertrain efficiency. The learned policy is coupled with a kinematics-based collision-avoidance filter to provide an explicit safety boundary during action execution. Experiments on NGSIM car-following events compare TD3 with human driving, an MPC-ACC controller, IDM, and three online reinforcement-learning baselines. Across three random seeds, TD3 achieved zero collision episodes, a mean TTC of 32.53s (SD 3 . 24 s), a mean headway of 1.31s (SD 0.17s), and the highest mean motor efficiency among the learned online controllers (85.52%, SD 0.62%). Direct energy accounting showed that TD3 achieved lower net energy intensity than DDPG, SAC, and PPO (112.57 Wh/km, SD 18.88 Wh/km). These results indicate that TD3 with a safety filter provides a practical compromise among safety, compact car-following, comfort, and powertrain efficiency, while also identifying direct battery-energy optimization as an important direction for future eco-driving rewards.
Authors
- Meijiao Wang
- He Tang (ORCID: https://orcid.org/0009-0008-3551-370X)
Institutions
- University of Shanghai for Science and Technology (CN)
Publication Details
- Journal
- Proceedings of the Institution of Mechanical Engineers Part D Journal of Automobile Engineering
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1177/09544070261505516
- Primary Topic
- Traffic control and management
- Type
- article
- Field-Weighted Citation Impact
- 0.00