Navigating Occluded Intersections with Autonomous Vehicles Using Deep Reinforcement Learning

Isele, David; Rahimi, Reza; Cosgun, Akansel; Subramanian, Kaushik; Fujimura, Kikuo · 2018 · OpenAlex-citations

DOI: 10.1109/icra.2018.8461233

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of navigating unsignaled intersections with autonomous vehicles, specifically focusing on scenarios involving occlusions where vehicle intent is obscured. The authors identify limitations in current state-of-the-art heuristic methods, particularly Time-to-Collision (TTC), which assume constant velocity, require full environmental knowledge, and often result in overly cautious behavior. To overcome these issues, the study explores Deep Reinforcement Learning (DRL) using Deep Q-Networks (DQNs) to learn policies that optimize safety, efficiency, and traffic flow without relying on hand-engineered rules. The experimental design utilizes the SUMO traffic simulator to evaluate DQN agents against a TTC baseline across five intersection scenarios: Right, Left, Left2 (crossing two lanes), Forward, and Challenge (high-density, six-lane intersection). The researchers tested three action representations: Sequential Actions (continuous acceleration/deceleration), Time-to-Go (discrete wait/go decisions), and Creep-and-Go (hybrid allowing slow forward movement). State representations included grid-based spatial data with real-valued heading, velocity, and TTC metrics. For occlusion experiments, the authors introduced visual obstructions and employed a dueling network architecture with prioritized replay. Performance was measured via success rate, collision percentage, average completion time, and average braking time induced in other traffic, with each method evaluated over 10,000 trials. Results demonstrate that DQN methods significantly outperform the TTC heuristic in efficiency and robustness. The DQN Time-to-Go agent achieved the highest success rates, reaching 98.46% in the difficult Challenge scenario compared to only 39.2% for TTC. DQN agents were substantially faster, with the Time-to-Go model completing tasks 28% faster than TTC on average. While DQN methods exhibited non-zero collision rates (e.g., 0.84% in the Challenge scenario), they remained comparable to TTC in terms of disrupting other traffic, as measured by average braking time. In occlusion experiments, the DQN agents successfully learned "creeping" behaviors—moving forward slowly to gather sensory information—allowing them to navigate safely despite limited visibility. However, the study notes that these learned policies had limited generalization capabilities when transferred to unseen scenarios without retraining. The significance of this work lies in demonstrating that DRL can surpass traditional heuristics in complex, partially observable driving environments. The findings highlight the potential for autonomous vehicles to learn active sensing behaviors, such as creeping, to mitigate occlusion risks. The paper concludes that while DQN offers superior performance in simulation, future research must address safety constraints to prevent overly aggressive learned behaviors and improve generalization to ensure reliable deployment in real-world settings.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness