dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models

2026-06-22

AI Summary

Discrete diffusion-based vision-language-action (dVLA) models are challenging to optimize via reinforcement learning due to the intractability of the marginal probability of the final action. This work addresses this issue by formulating the denoising process as a Markov decision process and proposing to optimize the joint probability of complete denoising trajectories. Furthermore, it introduces a task-adaptive dynamic scheduling mechanism for the number of denoising steps. As the first study to apply reinforcement learning to discrete diffusion VLA models, the proposed method achieves a 99.7% success rate on LIBERO and outperforms supervised fine-tuning baselines by 30.6% on RoboTwin 2.0, matching the performance of state-of-the-art world-action models.

Abstract

Vision-Language-Action (VLA) models have established a powerful paradigm for generalist robotic manipulation by grounding control into the semantic reasoning of VLMs. Prevailing architectures typically model actions continuously via diffusion or flow processes, or discretely through either autoregressive generation or parallel decoding. Recently, Discrete Diffusion VLAs (dVLAs) have emerged as a distinct alternative, unifying vision, language, and action into a single discrete token space via masked generative modeling. While combining iterative refinement with unified representations, its training has thus far been restricted to Supervised Fine-Tuning (SFT), leaving the potential of Reinforcement Learning (RL) for further policy refinement largely unexplored. A fundamental challenge in RL for dVLAs is that the marginal probability of the final action generated by dVLAs remains intractable. To solve this problem, we propose \textbf{dVLA-RL}, shifting the learning objective from the marginal action probability to the joint probability of the sampled generation path. Specifically, by modeling the denoising process as a Markov Decision Process (MDP), we mathematically formulate this path probability as a product of step-wise transitions. This trajectory-level objective provides a unified formulation that natively accommodates variable denoising steps. Leveraging this intrinsic fexibility, we introduce a unified step scheduling approach for complex multi-task learning, tailoring denoising steps to specific task complexities to maximize both success rates and computational effciency. Extensive evaluations demonstrate that our approach achieves a success rate of \textbf{99.7\%} on LIBERO. Furthermore, it establishes strong VLA-based results on RoboTwin 2.0 by delivering a \textbf{30.6\%} improvement over the SFT baseline, remaining competitive with strong World-Action Model baselines.

Problem

• Discrete Diffusion
• Vision-Language-Action Models
• Reinforcement Learning
• Marginal Probability
• Robotic Manipulation

Innovation

• Discrete Diffusion
• Reinforcement Learning
• Vision-Language-Action Models
• Trajectory-level Policy
• Markov Decision Process