FlowR2A Unifies Multimodal Driving Planning with Reward-to-Action Learning

Multimodal driving planning faces a fundamental challenge: balancing the benefits of dense reward supervision from scoring-based methods, which are limited to fixed action vocabularies, against the dynamic proposal generation capabilities of anchor-based methods, which suffer from sparse supervision. FlowR2A proposes a novel solution by reframing simulation-based rewards from discriminative targets into generative conditions. The core of FlowR2A involves learning the reward-conditioned action distribution using a flow-matching decoder. This approach allows the model to internalize the intricate correlations between an action and its various outcomes, encompassing safety, progress, comfort, and rule compliance. To effectively manage the trade-off between strict safety constraints and softer progress objectives, the method incorporates fine-grained per-timestep reward conditioning and reward noise augmentation. The generative formulation of FlowR2A naturally supports controllable test-time sampling through reward guidance and anchored sampling, leading to the production of high-quality driving proposals. The system has demonstrated state-of-the-art performance on both NAVSIM v1 and v2 benchmarks, delivering multimodal proposals that are significantly superior to those generated by previous methods.