EUVIMEDEUVIMED entdecken
Crossref

Safe and Efficient Autonomous Driving via Deep Reinforcement Learning with Safety-Constrained Soft Actor-Critic

Ramya Krishna Kurra, Sudhakar Murthy Molli

Research Digest on Engineering Management and Social Innovations · 10.03.2026

DOI: 10.46647/asz5hs29 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Research Digest on Engineering Management and Social Innovations, online erschienen am 10.03.2026. Verfasst von Ramya Krishna Kurra und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Autonomous driving is one of the most difficult and impactful uses of artificial intelligence, necessitating real-time decision-making in complicated, dynamical and safety-critical circumstances. The current paper introduces Safety-Constrained Soft Actor-Critic (SC-SAC) a new deep reinforcement learning framework that achieves optimal efficiency of driving, passenger comfort, and safety in diverse driving situations in urban and highway environments. Our method builds upon the entropy-regularized SAC algorithm that has a hierarchical safety constraint mechanism that imposes hard safety limits by formulating the Constrained Markov Decision Process (CMDP) and retaining the same sample efficiency and exploration advantages of maximum-entropy RL. We evaluate SC-SAC across four simulation environments totaling over 2,400 distinct driving scenarios against five established RL baselines—DQN, DDPG, PPO, TD3, and vanilla SAC. Experimental results demonstrate that SC-SAC achieves an average cumulative reward of 91.8, reducing safety violations by 73.9% compared to DQN and by 53.2% compared to PPO, while requiring only 4.1 million training steps—40.5% fewer than DQN. SC-SAC approaches human expert performance (93.2 average reward) while maintaining real-time inference capability at 45Hz. We also demonstrate ablation studies that isolate the role of each architectural component and are able to transfer them to the real world on a Tesla Model 3 instrumented research vehicle.

Zitieren

APA

Kurra, R. K., & Molli, S. M. (2026). Safe and Efficient Autonomous Driving via Deep Reinforcement Learning with Safety-Constrained Soft Actor-Critic. Research Digest on Engineering Management and Social Innovations. https://doi.org/10.46647/asz5hs29

Vancouver

Kurra RK, Molli SM. Safe and Efficient Autonomous Driving via Deep Reinforcement Learning with Safety-Constrained Soft Actor-Critic. Research Digest on Engineering Management and Social Innovations. 2026. doi:10.46647/asz5hs29

BibTeX
@article{Kurra2026,
  author = {Kurra, Ramya Krishna and Molli, Sudhakar Murthy},
  title = {Safe and Efficient Autonomous Driving via Deep Reinforcement Learning with Safety-Constrained Soft Actor-Critic},
  journal = {Research Digest on Engineering Management and Social Innovations},
  year = {2026},
  doi = {10.46647/asz5hs29}
}
Kostenlos starten →