Diffusion Policy Learning for Cooperative Ramp Merge Control from Vehicle–Infrastructure Collaborative Perception Data

Abstract

Cooperative ramp merging is still difficult due to the delayed, partially occluded and heterogeneous observations of connected vehicles. Develop a safety-shielded diffusion policy in this paper that converts vehicle-infrastructure collaborative perception into coordinated longitudinal control. Roadside and onboard tracks are aligned in time and associated to form a confidence-aware interaction graph. A conditional denoising process then samples joint acceleration sequences, and risk guidance biases the reverse trajectory towards acceptable gaps, while a lightweight safety projection repairs residual constraint violations. Optimize auxiliary demonstrations for training strategies and test various traffic demand, penetration rates, sensor noise, occlusion, and communication delay scenarios in a closed-loop micro-simulator. Compared with a rule-based cooperative controller, model predictive control, behavior cloning and an unshielded diffusion policy, the proposed method has reduced the average ramp delay by 24.7%, decreased speed variance by 31.2%, and increased successful merges to 98.1%. The 5th-percentile time-to-collision is 1.54 seconds for the strongest non-diffusion baseline and 2.06 seconds; only 3.8% of the executed actions modify the safety projection. Under 300ms communication delay, the success rate is 94.6%, and confidence conditioning and receding-horizon resampling avoid a sharp drop. Therefore, the above results show that diffusion-based multimodal action generation can effectively and stably support merge coordination when combined with explicit uncertainty representation and a minimally intrusive safety layer.