FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning

FlowR2A introduces a novel approach to multimodal driving planning by learning reward-conditioned action distributions through flow matching. Traditional methods in autonomous driving either rely on dense reward supervision with fixed action vocabularies or generate proposals dynamically from sparse ground-truth trajectories. These paradigms face fundamental limitations: the former is restricted in adaptability, while the latter struggles with proposal quality and diversity. FlowR2A bridges this gap by reframing simulation rewards as generative conditions, enabling the model to learn p(a|r) from dense trajectory-reward pairs. The architecture combines perception encoders extracting rich scene features, reward encoders translating multi-dimensional reward signals into conditioning vectors, and flow-based decoders that generate proposals via iterative denoising. During training, the model leverages fine-grained per-timestep reward signals and noise augmentation to balance safety constraints with progress objectives. At inference, reward guidance and anchored sampling provide controllable, diverse proposals, significantly improving proposal quality and safety metrics. Extensive evaluations on NAVSIM v1 and v2 benchmarks demonstrate that FlowR2A outperforms existing methods, achieving 92.8 and 88.9 PDMS respectively, with superior safety and efficiency scores. The approach not only advances the state of the art but also offers a flexible, interpretable framework for future autonomous driving systems. Its ability to generate high-quality, diverse proposals conditioned on complex reward signals marks a substantial step toward safer, more reliable, and more adaptable autonomous vehicles. Future work will focus on enhancing temporal consistency, reducing computational costs, and extending the framework to real-world multi-agent scenarios, promising a transformative impact on intelligent transportation.