World-Action Models (WAM) initialized from pre-trained video generation backbones have demonstrated remarkable potential for robot policy learning. However, existing approaches face two critical bottlenecks that hinder performance and deployment. First, jointly reasoning over future visual dynamics and corresponding actions incurs substantial inference overhead. Second, joint modeling often entangles visual and motion representations, making motion prediction accuracy heavily dependent on the quality of future video forecasts. To address these issues, we introduce GigaWorld-Policy, an action-centered WAM that learns 2D pixel-action dynamics while enabling efficient action decoding, with optional video generation. Specifically, we formulate policy training into two coupled components: the model predicts future action sequences conditioned on the current observation, and simultaneously generates future videos conditioned on the predicted actions and the same observation. The policy is supervised by both action prediction and video generation, providing richer learning signals and encouraging physically plausible actions through visual-dynamics constraints. With a causal design that prevents future-video tokens from influencing action tokens, explicit future-video generation is optional at inference time, allowing faster action prediction during deployment. To support this paradigm, we curate a diverse, large-scale robot dataset to pre-train an action-centered video generation model, which is then adapted as the backbone for robot policy learning. Experimental results on real-world robotic platforms show that GigaWorld-Policy runs 9x faster than the leading WAM baseline, Motus, while improving task success rates by 7%. Moreover, compared with pi-0.5, GigaWorld-Policy improves performance by 95% on RoboTwin 2.0.
Figures & tables
Figure 1: Comparison of GigaWorld-Policy with baselines on inference frequency and task success rate in real-world settings and on an A100 GPU.
Figure 2: (a) VLM-based VLA with auxiliary future supervision to densify training signals, but limited by the discriminative nature of VLMs ( Cen et al., 2025 ; Zhang et al., 2025 ; Ni et al., 2025a ) . (b) Joint action–video prediction with bidirectional attention ( Bi et al., 2025 ; Kim et al., 2026 ; Shen et al., 2025 ) , requiring future video generation at inference. (c) Two-stage design ( Pai et al., 2025 ; Li et al., 2026 ) that first generates future video and then gets actions via an Inverse Dynamics Model (IDM). This approach still requires full video prediction to obtain actions, incurring additional inference overhead. (d) GigaWorld-Policy: an action-centered WAM that leverages future visual dynamics as supervision during training, while making future-video prediction optional at inference for low-latency action generation.
Figure 3: Overview of GigaWorld-Policy, built on a pre-trained video generation backbone. During pre-training, the model learns action-relevant representations from large-scale videos. During post-training for policy learning, it jointly performs action-chunk prediction from the current observation and future video prediction as auxiliary supervision. At inference time, the future-video prediction branch is optional, enabling faster inference.
Figure 4: Attention mask for GigaWorld-Policy: action tokens Ta attend to states Ts and current observations To only, while future video tokens Tf also attend to actions.
Dataset
Hours
Dataset
Hours
Dataset
Hours
EgoDex ( Hoque et al., 2025 )
800
Agibot ( Bu et al., 2025 )
2,500
EGO4D ( Grauman et al., 2022 )
3,500
RoboMind ( Wu et al., 2024 )
300
RDT ( Liu et al., 2024b )
25
Open X-Embodiment ( O’Neill et al., 2024 )
3,500
DROID ( Khazatsky et al., 2024 )
350
ATARA ( Feng et al., 2025 )
10
Something-Something V2 ( Goyal et al., 2017 )
200
Table 1: Datasets and their estimated collection hours used in embodied data pre-training, covering egocentric human demonstrations and real-world robot manipulation videos.
Simulation Task
π0.5
X-VLA
Motus
Our
Clean
Rand.
Clean
Rand.
Clean
Rand.
Clean
Rand.
Adjust Bottle
0.79
0.83
1.00
0.99
0.89
0.93
1.00
1.00
Place A2b Left
0.62
0.60
0.48
0.49
0.88
0.79
0.94
0.88
Place Bread Skillet
0.38
0.46
0.77
0.67
0.86
0.83
0.94
0.90
Place Cans Plasticbox
0.40
0.47
0.97
0.98
0.98
0.94
1.00
1.00
Place Fan
0.25
0.36
0.80
0.75
0.91
0.87
0.92
0.94
Table 2: Evaluation on RoboTwin 2.0 Simulation. The best results are marked in bold , and the second-best results are underlined . GigaWorld-Policy achieves performance comparable to Motus while providing a 9× inference speedup.
Method
Time (ms)
SR in Simulation
SR in Real-World
Vision–Language–Action Models
π0.5 ( Intelligence et al., 2025 )
225
0.48
0.69
GigaBrain-0 ( Team et al., 2025b )
452
–
0.68
World–Action Models
Motus ( Bi et al., 2025 )
3231
0.88
0.76
Cosmos-Policy ( Kim et al., 2026 )
1413
–
0.58
Table 3: Inference latency on an NVIDIA A100, and Success Rate in simulation and real world. The best results are marked in bold , and the second-best results are underlined .
Method
Clean the Desk
Scan a QR Code
Sweep up Trash
Stack Bowls
Average
Vision–Language–Action Models
π0.5 ( Intelligence et al., 2025 )
0.75
0.55
0.65
0.80
0.69
GigaBrain-0 ( Team et al., 2025b )
0.70
0.65
0.60
0.75
0.68
World–Action Models
Motus ( Bi et al., 2025 )
0.80
0.75
0.70
0.80
0.76
Cosmos-Policy ( Kim et al., 2026 )
0.65
0.50
0.45
0.70
0.58
Table 4: Comparison of task success rates in real-world experiments. The best results are marked in bold , and the second-best results are underlined .
Figure 5: Real-world deployment of GigaWorld-Policy on PiPER arms for QR code scanning.
Figure 6: Real-world deployment of GigaWorld-Policy on PiPER arms for trash sweeping.
Δ
0
4
8
12
24
48
Success Rate ↑
0.60
0.76
0.78
0.83
0.80
0.76
Table 5: Effect of the sampling interval Δ on real-world task Success Rate. The model predicts K=⌊48/Δ⌋ future frames. K=0 indicates no future video prediction. The best results are marked in bold , and the second-best results are underlined .
Figure 12Table 13
Figure 9: Qualitative comparison between the self-attention baseline and our method. In the red boxes, our method more accurately predicts object state changes.
Simulation Task
π0.5
X-VLA
Motus
Our
Clean
Rand.
Clean
Rand.
Clean
Rand.
Clean
Rand.
Adjust Bottle
0.79
0.83
1.00
0.99
0.89
0.93
1.00
1.00
Beat Block Hammer
0.63
0.50
0.92
0.88
0.95
0.88
0.86
0.86
Blocks Ranking Rgb
0.43
0.35
0.83
0.83
0.99
0.97
0.92
0.96
Blocks Ranking Size
0.08
0.14
0.67
0.74
0.75
0.63
0.44
0.48
Click Alarmclock
0.97
0.93
0.99
0.99
1.00
1.00
1.00
1.00
Table 8: Evaluation on RoboTwin 2.0 Simulation. The best results are marked in bold , and the second-best results are underlined . GigaWorld-Policy achieves performance comparable to Motus while providing a 9× inference speedup.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Real-world deployment of GigaWorld-Policy on PiPER arms for stacking bowls and cleaning a desk.
World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead and hindering real-time closed-loop deployment. GigaWorld-Policy addresses this issue with an action-centered formulation, where future visual dynamics are used during training while action-only decoding is used at inference time. Building upon this framework, we present GigaWorld-Policy-0.5, an enhanced action-centered WAM designed for more efficient robot control. During pretraining, GigaWorld-Policy-0.5 adopts a mixed Action-Conditioned World Modeling (AC-WM) and WAM training strategy. This strengthens the coupling between visual dynamics and robot actions and improves the transferability of action representations for downstream policy learning. For efficient inference, GigaWorld-Policy-0.5 introduces a Mixture-of-Transformers architecture that separates visual dynamics modeling and action generation into specialized experts, reducing active computation during action-only inference and achieving 85 ms inference latency on a local RTX 4090 setup. In addition, we employ an agent-based AutoResearch pipeline to systematically search training configurations, enabling more efficient identification of optimal experimental setups while reducing the time and manual intervention required for hyperparameter tuning. Experiments and ablations show that GigaWorld-Policy-0.5 preserves the training benefits of future visual dynamics while improving inference efficiency for robot control.
World action models (WAMs) have emerged as a promising direction for robot policy learning, as they can leverage powerful video backbones to model the future states. However, existing approaches often rely on separate action modules, or use action representations that are not pixel-grounded, making it difficult to fully exploit the pretrained knowledge of video models and limiting transfer across viewpoints and environments. In this work, we present Action Images, a unified world action model that formulates policy learning as multiview video generation. Instead of encoding control as low-dimensional tokens, we translate 7-DoF robot actions into interpretable action images: multi-view action videos that are grounded in 2D pixels and explicitly track robot-arm motion. This pixel-grounded action representation allows the video backbone itself to act as a zero-shot policy, without a separate policy head or action module. Beyond control, the same unified model supports video-action joint generation, action-conditioned video generation, and action labeling under a shared representation. On RLBench and real-world evaluations, our model achieves the strongest zero-shot success rates and improves video-action joint generation quality over prior video-space world models, suggesting that interpretable action images are a promising route to policy learning.
World Action Models (WAMs) commonly rely on video generation to bridge visual world modeling and robot control. However, video-based WAMs face three coupled limitations: dense multi-frame future tokens make inference costly, full video prediction spends capacity on action-irrelevant temporal and appearance details, and long-horizon future imagination may introduce errors that mislead action prediction. These issues raise a simple question: Does world action model really need video generation? We propose ImageWAM, a simple WAM framework that repurposes pretrained image editing models for robot action prediction. In contrast to video generation, image editing provides a better-matched prior: it only needs to model a target-frame transformation, focuses on action-relevant current-to-target visual differences, and grounds task instructions to localized visual changes through edit pretraining. In practice, ImageWAM does not decode the target frame at inference time; instead, it conditions a flow-matching action expert on the KV caches produced by image-editing denoising, using them as a compact world-action context. ImageWAM outperforms standard VLA baselines and matching competitive WAMs without additional policy pretraining across different simulator and real-world experiments. It also reduces FLOPs to 1/6 and latency to 1/4 of video-based WAMs. Attention analysis further shows that editing caches focus on task-relevant change regions, supporting image editing as an effective alternative to video-based world-action modeling.
Yuyang Zhang, Wenyao Zhang, Zekun Qi +7
Shanghai Jiao Tong University · Tsinghua University · Tencent Robotics X +1