Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DriveFine: Refining-Augmented Masked Diffusion VLA for Precise and Robust Driving

Chenxu Dang, Sining Ang, Yongkang Li, Haochen Tian, Jie Wang, Guang Li, Hangjun Ye, Jie Ma, Long Chen, Yan Wang

TL;DR

This paper proposes DriveFine, a masked diffusion VLA model that combines flexible decoding with self-correction capabilities, and designs a novel plug-and-play block-MoE, which seamlessly injects a refinement expert on top of the generation expert.

Abstract

Vision-Language-Action (VLA) models for autonomous driving increasingly adopt generative planners trained with imitation learning followed by reinforcement learning. Diffusion-based planners suffer from modality alignment difficulties, low training efficiency, and limited generalization. Token-based planners are plagued by cumulative causal errors and irreversible decoding. In summary, the two dominant paradigms exhibit complementary strengths and weaknesses. In this paper, we propose DriveFine, a masked diffusion VLA model that combines flexible decoding with self-correction capabilities. In particular, we design a novel plug-and-play block-MoE, which seamlessly injects a refinement expert on top of the generation expert. By enabling explicit expert selection during inference and gradient blocking during training, the two experts are fully decoupled, preserving the foundational capabilities and generic patterns of the pretrained weights, which highlights the flexibility and extensibility of the block-MoE design. Furthermore, we design a hybrid reinforcement learning strategy that encourages effective exploration of refinement expert while maintaining training stability. Extensive experiments on NAVSIM v1, v2, and Navhard benchmarks demonstrate that DriveFine exhibits strong efficacy and robustness. The code will be released at https://github.com/MSunDYY/DriveFine.

DriveFine: Refining-Augmented Masked Diffusion VLA for Precise and Robust Driving

TL;DR

This paper proposes DriveFine, a masked diffusion VLA model that combines flexible decoding with self-correction capabilities, and designs a novel plug-and-play block-MoE, which seamlessly injects a refinement expert on top of the generation expert.

Abstract

Vision-Language-Action (VLA) models for autonomous driving increasingly adopt generative planners trained with imitation learning followed by reinforcement learning. Diffusion-based planners suffer from modality alignment difficulties, low training efficiency, and limited generalization. Token-based planners are plagued by cumulative causal errors and irreversible decoding. In summary, the two dominant paradigms exhibit complementary strengths and weaknesses. In this paper, we propose DriveFine, a masked diffusion VLA model that combines flexible decoding with self-correction capabilities. In particular, we design a novel plug-and-play block-MoE, which seamlessly injects a refinement expert on top of the generation expert. By enabling explicit expert selection during inference and gradient blocking during training, the two experts are fully decoupled, preserving the foundational capabilities and generic patterns of the pretrained weights, which highlights the flexibility and extensibility of the block-MoE design. Furthermore, we design a hybrid reinforcement learning strategy that encourages effective exploration of refinement expert while maintaining training stability. Extensive experiments on NAVSIM v1, v2, and Navhard benchmarks demonstrate that DriveFine exhibits strong efficacy and robustness. The code will be released at https://github.com/MSunDYY/DriveFine.
Paper Structure (28 sections, 7 equations, 7 figures, 7 tables)

This paper contains 28 sections, 7 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. End to end Autonomous Driving
  4. VLAs for Autonomous Driving
  5. Reinforcement Learning for Autonomous Driving
  6. Method
  7. Diffusion LLM for Planning
  8. Modatily Tokenization.
  9. Training and Inference.
  10. Block-level MoE for Refinement
  11. Training and Inference.
  12. Reinforcement Finetuning
  13. GRPO for Generation Expert.
  14. Hybrid RL for Refining Expert.
  15. Experiments
  16. ...and 13 more sections

Figures (7)

  • Figure 1: Comparison of decoding mechanisms for Action Tokens in Generative VLA Models. (a) Parallel refinement with multi-step denoising. (b) Token-by-token decoding. (c) Generate first in parallel, then refine.
  • Figure 2: PDMS-oriented RFT.
  • Figure 3: Failure cases caused by irreversible decoding in token-based VLAs.
  • Figure 4: Architecture Overview of DriveFine. Visual and textual inputs are jointly aligned into a unified language space. A set of masked tokens undergoes $s$ steps of parallel denoising followed by a single refinement step. The key difference between the generation expert and the refinement expert lies in their input tokens: the former decodes only the masked tokens, whereas the latter operates on unmasked tokens.
  • Figure 5: Reinforcement Fine-tuning Pipeline. The offline advantage and online advantage are jointly combined to form the hybrid advantage to supervise the training of the refinement expert.
  • ...and 2 more figures