DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching

Jiayi Chen; Wenxuan Song; Jiaxin Fang; Ruiqing Yin; Jingbo Wang; Shuai Chen; Jieyuan Pei; Yikai Qin

cs.ROMar 27, 2026

DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching

Authors: Jiayi Chen, Wenxuan Song, Jiaxin Fang, Ruiqing Yin, Jingbo Wang, Shuai Chen, Jieyuan Pei, Yikai Qin, +7 more

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · COCO Matrix · ShanghaiTech University · Shanghai Institute of Technical Physics, CAS · Tsinghua University, AIR

Abstract

Vision-Language-Action (VLA) models that encode actions using a discrete tokenization scheme have been widely adopted for robotic manipulation, but existing decoding paradigms remain fundamentally limited. Whether actions are decoded sequentially by autoregressive VLAs or in parallel by discrete diffusion VLAs, once a token is generated, it is typically fixed and cannot be revised in subsequent iterations. Consequently, early token errors cannot be effectively corrected later. We propose DFM-VLA, a discrete flow matching VLA that iteratively refines action tokens. DFM-VLA models a token-level probability velocity field that dynamically updates the full action sequence across refinement iterations. We investigate two approaches to constructing the velocity field: an auxiliary velocity-head formulation and an embedding-guided formulation. To further improve prediction accuracy, we introduce a metric-aligned action tokenizer (MAAT) tailored to the coarse-to-fine nature of DFM, together with a two-stage decoding strategy. Extensive experiments on CALVIN, LIBERO, LIBERO-Plus, and real-world manipulation tasks demonstrate the effectiveness of our approach. Our project is available at https://chris1220313648.github.io/DFM-VLA/.

Figures & tables

Figure 1: Comparison of decoding paradigms. 1) Autoregressive (AR) models require as many steps as the action sequence length, while 2) Discrete Diffusion (DD) models enable faster generation through parallel token updates. However, both have the same limitation that once an erroneous token is produced, it cannot be corrected in later iterations. We refer to this phenomenon as irreversible commitment . In contrast, 3) our Discrete Flow-matching (DFM) VLA performs full-sequence action refinement at every iteration, allowing token-level correction and improving action quality for robotic manipulation.

Figure 2: Overall architecture of DFM-VLA. Given language–vision context and noised action tokens $x_{t}$ , the model predicts clean actions $x_{1}$ and learns the velocity field via loss function $\mathcal{L}_{\text{ce}}$ or $\mathcal{L}_{\text{head}}$ .

Figure 3: Visualization of a single decoding step in the iterative refinement stage. After predicting final state $x^{\text{pred}}_{1}$ , the model does not directly output final action tokens. Instead, it constructs a velocity field to compute transition rates and selectively updates tokens to next state $x_{t+h}$ at each step.

Method	Tasks Completed in a Row					Avg. Len. $\uparrow$
Method	1	2	3	4	5	Avg. Len. $\uparrow$
RT-1 Brohan et al. (2023)	0.844	0.617	0.438	0.323	0.227	2.45
Robo-Flamingo ( Li et al., 2024 )	0.964	0.896	0.824	0.740	0.660	4.09
Deer ( Yue et al., 2024 )	0.982	0.902	0.821	0.759	0.670	4.13
GR-1 ( Wu et al., 2024 )	0.949	0.896	0.844	0.789	0.731	4.21
ReconVLA ( Song et al., 2025c )	0.980	0.900	0.845	0.785	0.705	4.23

Table 1: Comprehensive Evaluation of Long-Horizon Robotic Manipulation on the CALVIN Benchmark. w/o Embed denotes constructing the velocity field with the auxiliary head.

Method	LIBERO					LIBERO-Plus
Method	Spatial	Object	Goal	Long	Average	Camera	Robot	Language	Light	Background	Noise	Layout	Total
OpenVLA ( Kim et al., 2025b )	84.7	88.4	79.2	53.7	76.5	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
WorldVLA ( Cen et al., 2025 )	87.6	96.2	83.4	60.0	81.8	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0
$\pi_{0}$ -Fast ( Pertsch et al., 2025 )	96.4	96.8	88.6	60.2	85.5	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6
FlowVLA ( Zhong et al., 2025 )	93.2	95.0	91.6	72.6	88.1	51.0	25.4	70.2	77.6	79.4	58.3	47.0	56.2
DreamVLA ( Zhang et al., 2025b )	97.5	94.0	89.5	89.5	92.6	26.2	17.6	67.0	77.5	91.6	53.6	43.5	48.9

Table 2: Evaluation and comparison on the LIBERO and LIBERO-Plus benchmarks. All values are success rates (%).

Figure 4: Comparison of velocity field constructions across training steps on LIBERO.

Table 7

Figure 5: The real-world experimental setup and demonstrations of manipulation tasks.

Data Fraction	AR	DD	DFM (Ours)
10%	1.71	2.84	3.21
50%	3.01	3.88	4.03
100%	4.18	4.32	4.58

Table 5: Ablation on training data scale on CALVIN.

Task	$\pi_{0}$ -FAST	Dream-VLA	$\pi_{0.5}$	DFM-VLA w/o Embed	DFM-VLA
Grasp Moving Objects	50.0	57.5	82.5	70.0	75.0
Sort Vegetables	47.5	55.0	70.0	65.0	70.0
Place Veg. to Pot (Object Var.)	60.0	67.5	72.5	77.5	82.5
Pot Lift	50.0	57.5	65.0	70.0	77.5
Place Veg. to Pot (Pose Var.)	42.5	62.5	72.5	67.5	70.0
Place Block to Plate (Height Var.)	35.0	42.5	62.5	60.0	65.0

Table 6: Real-world success rates (%) across six tasks. Baselines include

\pi_{0}

-FAST ( Pertsch et al., 2025 ) , Dream-VLA ( Ye et al., 2025 ) , and

\pi_{0.5}

( Intelligence et al., 2025 ) .

Explore similar work

dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models
Jun 22, 2026Yuhao Wu, Yitian Liu, Weijie Shen +13Vision-Language-Action Models VLMs for Robotics
ScaleLab at SJTU · 1Shanghai Jiao Tong University · 3Baidu AI Cloud +3
SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance
Jun 29, 2026Tengyue Jiang, Chunpu Xu, Jiayue Kang +1Robotic Manipulation Vision-Language-Action Models
2East China University of Science and Technology · 3Hong Kong Polytechnic University · 4Xi’an University of Electronic Science and Technology +1
Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching
Sep 28, 2026Chenyu Zhang, Yuhang Cao, Daru Du +8Flow Matching Vision-Language-Action Models
IIIS, Tsinghua University · Shanghai Qizhi Institute

Jun 22, 2026J/K move · Enter open · S save

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

Yuhao Wu, Yitian Liu, Weijie Shen, Mishuo Han+12

ScaleLab at SJTU · 1Shanghai Jiao Tong University · 3Baidu AI Cloud+3

Vision-Language-Action Models VLMs for Robotics RL for VLMs

Open paper AskPDF arXiv

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 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 99.7% on LIBERO. Furthermore, it establishes strong VLA-based results on RoboTwin 2.0 by delivering a 30.6% improvement over the SFT baseline, remaining competitive with strong World-Action Model baselines.

Abstract

Explore similar work

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

SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance

Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching