FlowR2A is a multimodal driving planner that learns the reward-conditioned action distribution p(a | r) with flow matching. Instead of treating simulation rewards as discriminative targets, FlowR2A treats them as a condition, unifying the dense supervision of scoring-based methods with the generative proposal modeling of anchor-based methods. At inference, generation is steered toward high-reward trajectories via classifier-free guidance.</p>\n<p>Training pipeline:<br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/6451364c4b84eee0deea68fa/ZKrJ723K9gkPbwPg2QKHm.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6451364c4b84eee0deea68fa/ZKrJ723K9gkPbwPg2QKHm.png\" alt=\"pipeline-v2\"></a></p>\n<p>NAVSIM-v1 result:<br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/6451364c4b84eee0deea68fa/J1lAjOxNxibSGC9JKFmS5.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6451364c4b84eee0deea68fa/J1lAjOxNxibSGC9JKFmS5.png\" alt=\"navsim-v1\"></a></p>\n","updatedAt":"2026-06-24T03:34:28.520Z","author":{"_id":"6451364c4b84eee0deea68fa","avatarUrl":"/avatars/cc5c52f9b30f43b7faa47d0d5848a492.svg","fullname":"Xirui Li","name":"lixirui142","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":2,"identifiedLanguage":{"language":"en","probability":0.8296224474906921},"editors":["lixirui142"],"editorAvatarUrls":["/avatars/cc5c52f9b30f43b7faa47d0d5848a492.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.24231","authors":[{"_id":"6a3b4f210a86ac3098d5d729","name":"Xirui Li","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72a","name":"Zhe Liu","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72b","name":"Xiaoqing Ye","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72c","name":"Wenhua Han","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72d","name":"Yifeng Pan","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72e","name":"Junyu Han","hidden":false},{"_id":"6a3b4f210a86ac3098d5d72f","name":"Hengshuang Zhao","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/6451364c4b84eee0deea68fa/VUsTd2GGHQ5EC13EncCpA.mp4"],"publishedAt":"2026-06-23T00:00:00.000Z","submittedOnDailyAt":"2026-06-24T00:00:00.000Z","title":"FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning","submittedOnDailyBy":{"_id":"6451364c4b84eee0deea68fa","avatarUrl":"/avatars/cc5c52f9b30f43b7faa47d0d5848a492.svg","isPro":false,"fullname":"Xirui Li","user":"lixirui142","type":"user","name":"lixirui142"},"summary":"Multimodal driving planning faces a long-standing tension between two paradigms: scoring-based methods benefit from dense reward supervision but are confined to a fixed action vocabulary, while anchor-based methods generate proposals dynamically yet suffer from sparse supervision constrained to a single ground-truth trajectory. In this work, we propose FlowR2A, which resolves this tension by reframing simulation-based rewards from discriminative targets into generative conditions. By learning the reward-conditioned action distribution from dense trajectory-reward pairs with a flow-matching decoder, FlowR2A unifies the dense supervision of scoring-based methods with the proposal generation of anchor-based methods in a single generative model, forcing the model to internalize the correlation between an action and its outcomes in safety, progress, comfort, and rule compliance. To balance hard safety constraints against soft progress objectives, we introduce fine-grained per-timestep reward conditioning and reward noise augmentation. The generative formulation naturally supports controllable test-time sampling via reward guidance and anchored sampling, producing high-quality proposals. FlowR2A achieves state-of-the-art results on the NAVSIM v1 and v2 benchmarks, with multimodal proposals of substantially higher quality than prior methods.","upvotes":1,"discussionId":"6a3b4f220a86ac3098d5d730","projectPage":"https://lixirui142.github.io/flowr2a-ad/","githubRepo":"https://github.com/lixirui142/FlowR2A","githubRepoAddedBy":"user","ai_summary":"FlowR2A addresses the tension in multimodal driving planning by combining dense reward supervision with dynamic proposal generation through a flow-matching decoder that learns reward-conditioned action distributions.","ai_keywords":["flow-matching decoder","reward-conditioned action distribution","multimodal driving planning","simulation-based rewards","discriminative targets","generative conditions","dense trajectory-reward pairs","anchor-based methods","scoring-based methods","reward guidance","anchored sampling","NAVSIM v1","NAVSIM v2"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6451364c4b84eee0deea68fa","avatarUrl":"/avatars/cc5c52f9b30f43b7faa47d0d5848a492.svg","isPro":false,"fullname":"Xirui Li","user":"lixirui142","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.24231.md","query":{}}">
FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning
Abstract
FlowR2A addresses the tension in multimodal driving planning by combining dense reward supervision with dynamic proposal generation through a flow-matching decoder that learns reward-conditioned action distributions.
Multimodal driving planning faces a long-standing tension between two paradigms: scoring-based methods benefit from dense reward supervision but are confined to a fixed action vocabulary, while anchor-based methods generate proposals dynamically yet suffer from sparse supervision constrained to a single ground-truth trajectory. In this work, we propose FlowR2A, which resolves this tension by reframing simulation-based rewards from discriminative targets into generative conditions. By learning the reward-conditioned action distribution from dense trajectory-reward pairs with a flow-matching decoder, FlowR2A unifies the dense supervision of scoring-based methods with the proposal generation of anchor-based methods in a single generative model, forcing the model to internalize the correlation between an action and its outcomes in safety, progress, comfort, and rule compliance. To balance hard safety constraints against soft progress objectives, we introduce fine-grained per-timestep reward conditioning and reward noise augmentation. The generative formulation naturally supports controllable test-time sampling via reward guidance and anchored sampling, producing high-quality proposals. FlowR2A achieves state-of-the-art results on the NAVSIM v1 and v2 benchmarks, with multimodal proposals of substantially higher quality than prior methods.
Community
FlowR2A is a multimodal driving planner that learns the reward-conditioned action distribution p(a | r) with flow matching. Instead of treating simulation rewards as discriminative targets, FlowR2A treats them as a condition, unifying the dense supervision of scoring-based methods with the generative proposal modeling of anchor-based methods. At inference, generation is steered toward high-reward trajectories via classifier-free guidance.
Training pipeline:
NAVSIM-v1 result:
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.24231 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.24231 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.24231 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.