FlowR2A: Reward-to-Action model for multimodal driving planning

FlowR2A is a generative driving-planning model that learns the distribution of actions conditioned on simulation rewards, and it was published to arXiv (arXiv:2606.24231) on 23 Jun 2026. The paper, by Xirui Li, Zhe Liu, Xiaoqing Ye, Wenhua Han, Yifeng Pan, Junyu Han and Hengshuang Zhao, reports that FlowR2A achieves state-of-the-art results on the NAVSIM v1 and v2 benchmarks.

What is FlowR2A and how does it work?

FlowR2A learns a reward-conditioned action distribution from dense trajectory-reward pairs using a flow-matching decoder. The model reframes simulation-based rewards from discriminative targets into generative conditions, training on dense trajectory-reward pairs so the decoder internalizes how actions map to outcomes in safety, progress, comfort and rule compliance. The authors add fine-grained per-timestep reward conditioning and reward noise augmentation to balance hard safety constraints against softer progress objectives, and the generative formulation enables controllable test-time sampling via reward guidance and anchored sampling.

FlowR2A therefore unifies two previously competing paradigms by using dense reward supervision in a generative proposal model. The submitted paper file is 35,648 KB and the authors provide a project page linked from the submission.

How does FlowR2A compare to scoring-based and anchor-based approaches?

Scoring-based methods provide dense reward supervision but are confined to a fixed action vocabulary, while anchor-based methods generate proposals dynamically yet suffer from sparse supervision limited to a single ground-truth trajectory. FlowR2A resolves this tension by combining dense supervision with proposal generation in a single generative model. The paper describes FlowR2A as forcing the model to internalize the correlation between an action and its outcomes, while preserving the ability to generate multimodal proposals.

In concrete terms the paper contrasts the methods as follows: scoring-based approaches use dense rewards but a fixed action set; anchor-based approaches produce dynamic proposals but have sparse, single-trajectory supervision; FlowR2A learns from dense trajectory-reward pairs with a flow-matching decoder and supports anchored sampling and reward guidance at test time.

How was FlowR2A evaluated?

FlowR2A was evaluated on NAVSIM v1 and NAVSIM v2, where the authors report state-of-the-art results and that its multimodal proposals are of substantially higher quality than prior methods. The paper presents the model, training strategy and sampling mechanisms that produce those proposals, and it highlights the role of per-timestep reward conditioning and reward noise augmentation in balancing safety and progress objectives.

The submission appears on arXiv as version v1 with identifier arXiv:2606.24231 and a submission date of 23 Jun 2026.

Why it matters

FlowR2A changes the framing of reward supervision for planning from a discriminative target into a conditional generative signal. That matters because it lets a single model keep the dense feedback used to teach safety and route objectives while still producing diverse, test-time proposals. For applied driving planners this could reduce reliance on handcrafted action vocabularies or on training signals that only reflect one ground-truth trajectory.

What to watch

Watch for the project page and any released code or models linked from the arXiv entry, and for community evaluations on NAVSIM v1 and v2 that reproduce the paper's state-of-the-art claims. The next confirmation will be replication of the reported improvements in independent NAVSIM benchmarks or public code releases tied to the submission.