Diffusion-2BC: Hybrid Diffusion and Regression Training for Offline Behavior Cloning in Autonomous Driving

This research proposes Diffusion-2BC, which integrates a diffusion denoising objective with an auxiliary deterministic behavior-cloning loss over a shared visual encoder. Evaluated in controlled and CARLA environments, the method reduces mean mask-distance error by 10% compared to a diffusion baseline and 85% compared to standard behavior cloning, showing improved closed-loop reliability and multimodal action prediction.