cs.ROMar 26, 2026
SaveMMaDA-VLA: Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation
Organizations: Westlake University Hangzhou, China · Zhejiang University Hangzhou, China · East China University of Science and Technology Shanghai, China · Huawei Technologies Ltd. Hangzhou, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China
Abstract
Vision-Language-Action (VLA) models map visual observations and natural-language instructions to robot actions; however, hierarchical and autoregressive paradigms often incur architectural overhead, accumulate long-horizon errors, and require auxiliary modules to capture environment dynamics. To this end, we present MMaDA-VLA, a fully native, pretrained discrete diffusion VLA that unifies multi-modal understanding and generation. Specifically, MMaDA-VLA uses a shared discrete token space to jointly denoise a future goal observation and an action chunk, grounding actions in predicted visual outcomes without an auxiliary world model. In this way, parallel, order-free refinement improves long-horizon consistency. Extensive experiments and comprehensive analyses demonstrate that MMaDA-VLA achieves an average success rate of 98.0% on LIBERO and an average successful sequence length of 4.78 on CALVIN, while performing strongly in real-world settings. The project page is available at https://yliu-cs.github.io/MMaDA-VLA.
Figures & tables
Figure 1. (a) Comparison of conventional VLA methods and MMaDA-VLA . (b) Comparison with discrete diffusion methods. (c) Performance comparison with state-of-the-art methods. Cur., Obs., Gen., and DDVLA denote current, observation, generation, and Discrete Diffusion VLA, respectively. Three panels compare vision-language-action approaches. The left panel contrasts policy-head, world-model, unified-VLA, and MMaDA-VLA pipelines. The middle panel shows which competing models use autoregressive or discrete-diffusion backbones, VLA pre-training, fine-tuning, and inference. The right panel reports MMaDA-VLA at 98.0 percent on LIBERO, 4.78 average length on CALVIN, and 89.2 percent on real-world tasks.
Figure 2. Schematic of the MMaDA-VLA framework and training pipeline. MMaDA-VLA combines current observations, task instructions, masked goal observations, and masked action chunks in a single large diffusion vision-language-action model. Iterative parallel denoising predicts a future goal observation and a chunk of seven-dimensional robot actions, followed by closed-loop execution. The model is first pre-trained on large-scale robot data and then fine-tuned on downstream tasks.
| Method | LIBERO | Method | CALVIN | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Spatial | Object | Goal | Long | Avg. | 1/5 | 2/5 | 3/5 | 4/5 | 5/5 | Avg. Len. | ||
| Diffusion Policy ( Chi and others, 2023 ) | 78.3 | 92.5 | 68.3 | 50.5 | 72.4 | Diffusion Policy ( Chi and others, 2023 ) | 40.2 | 12.3 | 2.6 | 0.8 | 0.0 | 0.56 |
| Octo ( Ghosh and others, 2024 ) | 78.9 | 85.7 | 84.6 | 51.1 | 75.1 | RT-1 ( Brohan and others, 2023 ) | 53.3 | 22.2 | 9.4 | 3.8 | 1.3 | 0.90 |
| OpenVLA ( Kim and others, 2024 ) | 84.9 | 88.4 | 79.2 | 53.7 | 76.5 | VLAS ( Zhao and others, 2025b ) | 87.2 | 64.2 | 40.9 | 28.1 | 19.6 | 2.40 |
| SpatialVLA ( Qu and others, 2025 ) | 88.2 | 89.9 | 78.6 | 55.5 | 78.1 | RoboFlamingo ( Li and others, 2024b ) | 82.4 | 61.9 | 46.6 | 33.1 | 23.5 | 2.47 |
| DiT Policy ( Hou and others, 2024 ) | 84.2 | 96.3 | 85.4 | 63.8 | 82.4 | Deer ( Yue and others, 2024 ) | 84.8 | 72.3 | 54.9 | 44.6 | 33.5 | 2.90 |
Table 1. Performance comparison on LIBERO ( Liu and others, 2023a ) and CALVIN ( Mees et al., 2022 ) . For LIBERO, we report task success rates (%); for CALVIN, we report subtask success rates (%) and average sequence length. Continuous-action methods are shown in grey .
Figure 3. Real-world experiment setup, task examples, and performance comparison. The experiment uses an AgileX PiPER arm with wrist-view and third-person cameras and diverse tabletop objects. Four task columns show pick-and-place, block stacking, drawer storage, and bowl-and-cup organization. Bar charts compare MMaDA-VLA with GR00T N1.6; MMaDA-VLA reaches 86.7, 93.3, 93.3, and 83.3 percent success, respectively, versus 63.3, 70.0, 66.7, and 56.7 percent.
| Dataset | Ratio (%) | Samples | Dataset | Ratio (%) | Samples |
|---|---|---|---|---|---|
| DROID ( Khazatsky and others, 2024 ) | 49.94 | 26675496 | Austin Sirius Dataset ( Liu and others, 2023b ) | 0.52 | 277701 |
| BC-Z ( Jang and others, 2021 ) | 9.95 | 5314594 | Taco Play ( Rosete-Beas and others, 2022 ; Mees and others, 2023 ) | 0.38 | 201002 |
| Language Table ( Lynch et al., 2023 ) | 9.88 | 5276570 | Roboturk ( Mandlekar and others, 2018 ) | 0.30 | 161268 |
| Furniture Bench Dataset ( Heo and others, 2023 ) | 7.35 | 3927655 | IAMLab CMU Pickup Insert ( Saxena and others, 2023 ) | 0.27 | 143715 |
| Fractal ( Brohan and others, 2023 ) | 6.44 | 3438450 | Berkeley Autolab UR5 ( Chen and others, ) | 0.16 | 84197 |
| Bridge V2 ( Walke and others, 2023 ) | 3.15 | 1680298 | Viola ( Zhu and others, 2022b ) | 0.13 | 68371 |
Table 2. Dataset mixture details for MMaDA-VLA .
| Configuration | Pre-Training | Fine-Tuning | Configuration | Pre-Training | Fine-Tuning |
|---|---|---|---|---|---|
| Backbone | MMaDA-8B-Base ( Yang and others, 2025 ) | Image Tokenizer | MAGVIT-v2 ( Yu et al., 2024 ) | ||
| Global Batch Size | 640 | 80 ♯ & 320 † | Numerical Precision | BFloat16 | |
| Optimizer | AdamW ( Loshchilov and Hutter, 2019 ) | Learning Rate | |||
| Epochs | 1 | 20 40 ♯ & 2 † | Weight Decay | 0.01 | |
| Learning-Rate Schedule | Cosine Decay | Warm-up Ratio | 0.01 | ||
| Textual Instruction Length | 128 | Action Chunk Size | 5 | 5 ♯ & 10 † | |
Table 3. Hyperparameter settings. ♯ and † denote the settings used for LIBERO and CALVIN, respectively.
Figure 4. MMaDA-VLA execution examples on real-world tasks. Time-ordered image sequences show MMaDA-VLA executing four real-world tasks: placing a banana into a bowl, stacking colored blocks, opening a drawer to store an object and then closing it, and organizing bowls and cups on a table.
| Method | Average Length |
|---|---|
| MMaDA-VLA (w/o Pre-Training) | 4.56 |
| w/o World-Model | 4.08 |
| w/o Parallel Denoising | 4.38 |
| w/ Causal Attention | 4.49 |
| w/ Bidirectional Attention | 4.52 |
Table 4. Ablation studies on CALVIN.
| Method | LIBERO | CALVIN |
|---|---|---|
| Average Success Rate (%) | Average Sequence Length | |
| MMaDA-VLA | 98.0 | 4.78 |
| w/o Pre-Training | 94.5 | 4.56 |
Table 5. Effectiveness of pre-training.
Figure 5. Visualization of generated goal observations versus ground-truth observations. For each timestep, we form a single composite frame by vertically concatenating the third-person view (top) and the wrist-mounted view (bottom). Two task sequences compare ground-truth observations in the top row with generated observations in the bottom row. Across the tasks of putting a box and butter in a basket and putting a bowl in a drawer before closing it, the generated sequences preserve the main object motion and task progression while losing some fine visual detail.
Figure 6. Pre-training visualization. The left plot shows cross-entropy loss varying within a bounded range over roughly 110,000 pre-training steps. The right plot shows the sampled mask-rate density over values from zero to one.
Figure 7. Inference latency and performance under different iterative denoising and cache-refresh settings. A scatter plot compares average CALVIN task length against inference latency under different iterative-denoising and cache-refresh settings. The best trade-off region is around 1.4 to 1.5 seconds of latency with an average length above 4.55; 24 iterative steps achieve the highest average length of about 4.65.