dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models
2026-06-22 • Robotics
Robotics
AI summaryⓘ
The authors focus on improving Vision-Language-Action (VLA) models that help robots understand and perform tasks by linking words, images, and actions. They introduce a new training method called dVLA-RL, which uses reinforcement learning to refine how the model predicts sequences of actions. Instead of trying to calculate the overall chance of an action, they treat the step-by-step action generation as a process and optimize it accordingly. Their approach allows for adapting the number of steps taken depending on task difficulty, leading to better success rates and efficiency. Experiments show that their method significantly improves performance on robotic task benchmarks compared to previous models.
Vision-Language-Action (VLA) modelsDiscrete DiffusionReinforcement Learning (RL)Masked generative modelingMarkov Decision Process (MDP)Trajectory-level objectiveStep schedulingMulti-task learningRobotic manipulationPolicy refinement
Authors
Yuhao Wu, Yitian Liu, Weijie Shen, Mishuo Han, Wenjie Xu, Haotian Liang, Zhongshan Liu, Yinan Mao, Lei Xu, Xinping Guan, Ru Ying, Ran Zheng, Wei Sui, Xiaokang Yang, Wenbo Ding, Yao Mu
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.