Roundabouts present a major challenge for automated driving because human behavior is heterogeneous and non-deterministic, driving intentions are unknown, and interaction complexity is high. These factors create uncertainty about whether the conflict zone will be blocked or available at the moment of entry. According to a paper published on arXiv, researchers have developed ROSA-RL (Roundabout Optimized Speed Advisory with Reinforcement Learning) to address this problem.
Probabilistic Conflict Forecasting with Transformers
ROSA-RL employs a Transformer-based model to predict conflict zone occupancy over a five-second horizon. The model captures multi-agent interactions, enabling it to anticipate upcoming conflicts and available gaps. According to the paper, the prediction outputs encode uncertainty in future motion and intent, which is then used to augment the state of a classical reinforcement learning (RL) framework. This allows the system to coordinate speed in an uncertainty-aware manner.
Uncertainty-Aware Reinforcement Learning
The core innovation of ROSA-RL is its ability to handle uncertainty explicitly. By incorporating probabilistic conflict forecasts into the RL state representation, the system can make safer and more efficient decisions in mixed traffic environments where human-driven vehicles and automated vehicles interact. The researchers note that this approach closes the gap to an ideal setting that assumes fully known occupancy, while improving both traffic efficiency and safety.
Simulation Evaluation and Results
ROSA-RL was evaluated in simulations grounded in real-world data. According to the paper, the system effectively handles uncertainty and outperforms a comparable model-based baseline. The results demonstrate that the uncertainty-aware RL framework can nearly match the performance of an ideal system with complete knowledge of future occupancy, without requiring that perfect information.
The source code for ROSA-RL is publicly available, as noted in the paper.
Implications for Mixed Traffic Automation
While ROSA-RL is specifically designed for roundabouts, the underlying approach—combining Transformer-based multi-agent prediction with uncertainty-aware reinforcement learning—could be extended to other traffic scenarios involving high interaction complexity and uncertain human behavior. The paper, authored by Schlamp, Anna-Lena; Gerner, Jeremias; Bogenberger, Klaus; Huber, Werner; and Schmidtner, Stefanie, is listed under Computer Science > Artificial Intelligence on arXiv.
