Rollout-Based Decision-Making for Autonomous Driving Left Turn at Unsignalized Intersections
- 23 February 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (4) , 4438-4447
Abstract
Unsignalized intersections pose significant safety and efficiency challenges for autonomous driving due to ambiguous right-of-way and complex, multi-directional vehicle interactions within overlapping conflict zones. In this study, we propose a Rollout-Enhanced Attention (REA) framework, which synergizes a rollout algorithm with a multi-head cross-attention mechanism to optimize driving decisions. Specifically, the multi-head cross-attention mechanism effectively extracts interaction features between the ego vehicle and surrounding vehicles, prioritizing those with higher collision risks. Incorporating these features, the rollout algorithm executes online multi-step lookahead planning by leveraging the action-value estimates from a Deep Q-Network (DQN), thereby optimizing action selection. Extensive experiments in unsignalized intersection scenarios with varying traffic densities demonstrate that the REA model outperforms standard DQN and other baselines. Furthermore, comparisons between different rollout variations highlight the framework’s advantages in safety, efficiency, and decision-making accuracy.Keywords
Funding Information
- National Natural Science Foundation of China (62303297)
- Shanghai Sailing Program (23YF1413100)
- National Center for Science and Engineering Statistics
This publication has 22 references indexed in Scilit:
- Uncertainty-aware Reinforcement Learning for Autonomous Driving with Multimodal Digital Driver Guidance2024
- Multiagent Rollout with Reshuffling for Warehouse Robots Path PlanningIFAC-PapersOnLine, 2023
- Decision Making through Occluded Intersections for Autonomous Driving2019
- Navigating Occluded Intersections with Autonomous Vehicles Using Deep Reinforcement Learning2018
- A random parameters probit model of urban and rural intersection crashesAccident Analysis & Prevention, 2015
- Human-level control through deep reinforcement learningNature, 2015
- Provably safe and robust learning-based model predictive controlAutomatica, 2013
- Congested traffic states in empirical observations and microscopic simulationsPhysical Review E, 2000
- Rollout Algorithms for Stochastic Scheduling ProblemsJournal of Heuristics, 1999
- Rollout Algorithms for Combinatorial OptimizationJournal of Heuristics, 1997