Vision-language-action models (VLAs) inherit strong semantic grounding from pretrained vision-language backbones but are typically optimized for predicting actions rather than future observations. They can see and act, but they do not imagine the future before acting. World action models (WAMs) built on video diffusion backbones can imagine but treat language as frozen conditioning on a continuous latent space. Unified models bring these modalities into one architecture, but they either decode autoregressively, one token at a time, or keep video continuous with an auxiliary action head. In this work, we present SUAVE, a Single vocabulary Unified Action-Video modEl in which a masked diffusion transformer generates video and actions conditioned on language, with all three modalities represented as discrete tokens in a shared sequence. Choosing which tokens to mask at inference turns the same network into a world model, a robot policy, or a video-action model. For action-free co-training, the action positions of unlabeled video are filled with mask tokens and excluded from the loss. Simulation and real-world experiments demonstrate two findings. First, a single SUAVE model predicts long-horizon video and acts as a policy, competitive with dedicated world models and specialized action policies on static and dynamic manipulation tasks. On a real robot, our model generates subgoal images and an action chunk spanning one second of motion in 1,030 ms on an RTX 5090 GPU, sustaining closed-loop control at 2.5 actions per second. Second, pretraining on robot video and co-training on human video substantially improves policy performance and zero-shot robustness to distribution shift. Together, these results show that masked diffusion is a practical and versatile foundation for unified video-action modeling.
Figures & tables
Fig. 2: Architecture. Left: The instruction is tokenized by the LLaDA tokenizer, and the video context is encoded by a MAGViT-v2 tokenizer. The goal and action tokens start fully masked, and a masked diffusion transformer, initialized from LLaDA with MMaDA’s image-token embeddings plus a new action vocabulary, denoises them into the imagined goal frame and the action chunk. Middle: Each of the 32 layers is a standard LLaDA block whose self-attention uses a block-causal mask. Right: Each cell in the mask is one block. Attention is bidirectional within a block and causal across blocks. The clean prefix is computed once and cached, and only the masked suffix is recomputed at each denoising step.
Fig. 3: Real robot task score under nominal conditions and distribution shift. Each violin is the posterior of the mean Task Score for one model variant on one task, computed by a Bayesian bootstrap over its 24 rollouts. The black bar is the mean. The left half shows nominal conditions and the right half the distribution shift with added clutter, unseen shoes, altered lighting. Models labeled with different letters are statistically distinguishable.
Fig. 4: Qualitative Comparison of Long-Horizon Video Prediction. Rows show the ground-truth rollout (GT), SUAVE (red), and four baselines. Left: predicted frames for the Sweep Pile task at the indicated times (in seconds). Right: final predicted frame for Pick Vegetables and Organize Shoes. White circles mark failure modes.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FVD ↓
Cosmos [ 11 ]
17.8 ± 1.2
0.61 ± 0.04
0.22 ± 0.04
472 ± 33
iVideoGPT [ 9 ]
17.2 ± 2.3
0.39 ± 0.12
0.46 ± 0.15
2939 ± 1025
UVA [ 13 ]
19.3 ± 1.7
0.60 ± 0.06
0.33 ± 0.04
2368 ± 1356
UWM [ 12 ]
15.9 ± 1.0
0.38 ± 0.05
0.42 ± 0.09
1922 ± 524
UniVLA [ 14 ]
17.9 ± 1.9
0.58 ± 0.05
0.27 ± 0.06
1095 ± 195
SUAVE-Scratch
18.4 ± 1.4
0.55 ± 0.04
0.23 ± 0.05
1883 ± 588
TABLE I: Video prediction on real tasks. Rollouts from the first observed frame to the end of the episode on 8 held-out episodes per task. All predicted frames of a task are pooled against the raw 256×256 camera frame. Baselines are trained by us with the authors’ released recipes and weights on the same demonstrations. SUAVE rows are the same checkpoints of Fig. 3 .
Method
Spatial
Object
Goal
Long
Avg
DP [ 30 ]
78.3 ± 1.1
92.5 ± 0.7
68.3 ± 1.2
50.5 ± 1.3
72.4 ± 0.7
OpenVLA [ 1 ]
84.7 ± 0.9
88.4 ± 0.8
79.2 ± 1.0
53.7 ± 1.3
76.5 ± 0.6
CoT-VLA [ 5 ]
87.5 ± 1.4
91.6 ± 0.5
87.6 ± 0.6
69.0 ± 0.8
81.1 ± 0.6
UWM ∗ [ 12 ]
84.0 ± 1.6
84.4 ± 1.6
87.4 ± 1.5
84.0 ± 1.6
85.0 ± 0.8
UniVLA [ 14 ]
94.2 ± 1.0
97.2 ± 0.7
92.4 ± 1.2
91.8 ± 1.2
93.9 ± 0.5
π0.5 [ 3 ]
98.6 ± 0.5
99.0 ± 0.4
98.6 ± 0.5
93.2 ± 1.1
97.4 ± 0.4
TABLE II: LIBERO success rates (%, higher is better). Mean and standard error over 500 trials per suite. ∗ UWM is retrained by us as a single multitask language-conditioned policy. Bold marks the best per column.
Long horizon suite
Method
Pretraining corpus
LIBERO
LIBERO-Plus
(zero-shot)
SUAVE-Scratch
none
87.7 ± 0.7
52.1 ± 1.0
SUAVE-Pretrain
DROID
87.2 ± 0.7
53.1 ± 1.0
SUAVE-Cotrain
DROID + SSv2
88.7 ± 0.7
53.9 ± 1.0
SUAVE-Pretrain
LIBERO-90
93.6 ± 0.5
63.4 ± 1.0
TABLE III: Pretraining ladder on LIBERO-Long (%): All rows finetune the same model on LIBERO-Long and differ only in pretraining corpus. We show the original LIBERO mean and standard error over 4 seeds × 500 trials each and the LIBERO-Plus zero-shot success on 2,519 trials each.
Method
Camera
Robot
Language
Light
Background
Noise
Layout
Total
Zero-shot: trained on LIBERO demonstrations only
OpenVLA [ 1 ]
0.8
3.5
23.0
8.1
34.8
15.2
28.5
15.6
OpenVLA-OFT [ 7 ]
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
OpenVLA-OFT w
10.4
38.7
70.5
76.8
93.6
49.9
69.9
55.8
OpenVLA-OFT m
55.6
21.7
81.0
92.7
91.0
78.6
68.7
67.9
π0 [ 2 ]
13.8
6.0
58.8
85.0
81.4
79.0
68.9
53.6
TABLE IV: LIBERO-Plus success rates (%) by perturbation category. All rows cover the full benchmark (10,030 tasks, one trial each). SUAVE-Cotrain is pretrained on LIBERO-90 and SSv2. Bold marks the best per column.
Level 1
Level 2
Level 3
constant velocity
polynomial
stochastic, abrupt
Method
SR
MS
RC
SR
MS
RC
SR
MS
RC
UWM [ 12 ]
5.9 ± 0.5
20.7 ± 0.4
28.5 ± 0.5
5.3 ± 0.5
20.4 ± 0.4
29.0 ± 0.5
3.6 ± 0.4
17.3 ± 0.4
26.4 ± 0.4
π0 [ 2 ]
3.6 ± 0.4
13.8 ± 0.4
21.6 ± 0.4
2.4 ± 0.3
12.1 ± 0.3
20.0 ± 0.4
0.8 ± 0.2
10.4 ± 0.3
18.3 ± 0.4
π0 -FAST [ 8 ]
10.8 ± 0.6
22.7 ± 0.5
30.3 ± 0.6
8.3 ± 0.6
19.9 ± 0.5
27.9 ± 0.5
2.6 ± 0.3
15.1 ± 0.4
24.8 ± 0.4
π0.5 [ 3 ]
6.2 ± 0.5
16.1 ± 0.5
23.4 ± 0.5
4.2 ± 0.4
13.9 ± 0.4
21.7 ± 0.4
1.7 ± 0.3
11.6 ± 0.3
19.8 ± 0.4
TABLE V: DOMINO across dynamics levels (%): All rows are single checkpoints trained on the same Franka demonstrations and evaluated on the 25-task subset at 100 episodes per task. We report SR, MS, and RC as mean and standard error over 2,500 episodes per level. Bold marks the best per column.
Vision-Language-Action (VLA) models generalize semantically well but often lack fine-grained modeling of world dynamics. We present Motubrain, a unified World Action Model that jointly models video and action under a UniDiffuser formulation with a three-stream Mixture-of-Transformers architecture. A single model supports policy learning, world modeling, video generation, inverse dynamics, and joint video-action prediction, while scaling to heterogeneous multimodal data such as video-only, task-agnostic, and cross-embodiment robot data. Building on Motus, Motubrain further introduces unified multiview modeling, an independent text stream for stronger language-action coupling, a shared cross-embodiment action representation, and an efficient post-training and deployment recipe for long-horizon real-world control. Our inference stack combines step reduction, compilation, FP8 quantization, DiT caching, V2A-style action-only inference, and real-time chunked closed-loop execution, achieving over 50x speedup over a naive baseline and up to 11 Hz inference. Experimentally, Motubrain achieves 95.8% and 96.1% average success on RoboTwin 2.0 under clean and randomized settings, respectively, attains the strongest reported EWMScore in our WorldArena comparison, and adapts to new humanoid embodiments with only 50--100 trajectories. These results show that unified world action models can scale in generality, predictive accuracy, and real-world deployability.
Vision-language-action (VLA) models aim to understand natural-language instructions and visual observations, and to generate and execute corresponding actions as embodied agents. Recently, autoregressive token-based action generation has driven the development of many representative VLA models. However, this paradigm often reduces action generation to next-token prediction, thereby lacking explicit modeling of the spatiotemporal structure of action sequences and the disentanglement between vision-language representations and actions, which can limit performance in long-horizon and complex scenarios. In this paper, we propose TS-Mask VLA, a vision-language-action framework for robot manipulation. TS-Mask VLA is built upon two key designs: (1) a Discrete Diffusion Action Expert equipped with a Bridge Attention conditioning bridge, which enables multi-layer conditioning from the VLM and facilitates more accurate and stable action generation; and (2) a temporal-spatial 2D masking strategy for discrete action tokens that strengthens the model's understanding of cross-time dependencies and inter-dimensional coupling, leading to more structurally consistent action sequences. We conduct extensive experiments on simulation benchmarks and real-world tasks. On LIBERO, TS-Mask VLA achieves a 95.7 percent average success rate with only 0.5B parameters, outperforming significantly larger models. On CALVIN, it attains the best average sequence length of 4.19 and strong long-horizon performance. Comprehensive analyses and ablations further validate the effectiveness of our design.
Shengzhuo Yang, Ronghao Yu, Chuanjie Lv +5
Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China
Many recent robot policies pursue stronger control by using large pretrained vision-language models (VLMs) as the action backbone. We introduce WorldDiT, a unified diffusion transformer architecture that couples action generation with visual world modeling and achieves strong performance without a large pretrained VLM action backbone. During training, a single diffusion transformer generates continuous action chunks and predicts normalized RGB patch targets from future camera frames. Across four LIBERO simulation suites, WorldDiT lies on the reported Pareto frontier for total model parameters and mean success among methods reporting all four suites. These results provide a strong sub-billion-parameter baseline for future scaling studies.