World Action Models (WAMs) leverage pretrained video generative models for robotic control by jointly predicting future visual observations and actions. The visual prediction objective plays a key role in adapting pretrained video representations to the target robot domain. Yet standard visual prediction inherited from video pretraining treats all visual regions uniformly, despite their differing relevance to robotic control. We therefore investigate the spatial allocation of visual supervision in WAMs and show that focusing supervision on dynamic regions can substantially improve policy performance. Building on this finding, we introduce Motion-Guided Asymmetric Prediction (MoGAP), which augments WAM training with motion-focused visual prediction. MoGAP allocates noise asymmetrically across regions, training the model to reconstruct more heavily corrupted dynamic regions using less corrupted static regions as context. MoGAP is trained alongside standard video-action prediction through a training-only auxiliary prediction path within the shared video diffusion backbone. It uses motion cues derived directly from training videos, requiring neither an external teacher nor additional modalities. Across robot manipulation benchmarks and WAM architectures, MoGAP consistently improves policy performance, increasing success rates on the GR-1 tabletop benchmark from 38.8% to 47.5% for Fast-WAM (Wan2.2-5B) and from 52.3% to 62.3% for Cosmos3-Nano.
Figures & tables
Figure 1: Overview. (a) We spatially reallocate visual supervision in World Action Models (WAMs) toward dynamic regions. This reallocation emphasizes robot motion and object interactions while downweighting static, task-irrelevant visual details, such as complex tabletop textures. (b) Our method, Motion-Guided Asymmetric Prediction (MoGAP), substantially improves policy performance across different WAM setups on GR-1 Tabletop, a challenging humanoid manipulation benchmark.
Figure 2: Motion-Guided Asymmetric Prediction (MoGAP). MoGAP is a visual prediction task that promotes control-relevant representation learning in World Action Models. MoGAP introduces regional asymmetry in noise levels, guided by temporal differences between video latents: dynamic regions receive higher noise levels, while less corrupted static regions serve as context for prediction.
Figure 4: Attention mask of MoGAP . The MoGAP tokens ztτM are omitted at inference time.
Method
PnP
Articulated
24 Tasks Average
VLA-based methods
GR00T-N1.6 ( Bjorck et al., 2025 )
55.4
24.2
47.6
StarVLA- α ( Ye et al., 2026a )
58.4
54.0
57.3
RLDX-1 ( Kim et al., 2026a )
59.0
57.7
58.7
WAM-based methods
Fast-WAM ( Yuan et al., 2026 )
33.8
53.7
38.8
Table 1: Comparison with state-of-the-art VLAs & WAMs. Results include the success rate (%) of representative VLAs and WAMs on GR-1 tabletop. Results of Fast-WAM and Cosmos3 are reproduced while others are from the original work. Best results are highlighted in bold .
Method
Teacher
24 Tasks Average
Cosmos3
–
52.3
w/ Geometry Forcing ( Wu et al., 2026 )
VGGT
53.9
w/ AGRA ( Qiu et al., 2026 )
DINOv2
55.4
w/ TSA ( Kim et al., 2026b )
VideoMAEv2
56.8
w/ VideoREPA ( Zhang et al., 2025 )
VideoMAEv2
57.8
w/ VideoREPA ( Zhang et al., 2025 )
V-JEPA2.1-L
59.2
Table 2: Comparison to teacher-based supervision. Results report the average success rate (%) on GR-1 tabletop, with each method applied to the same baseline WAM, Cosmos3-Nano ( Agarwal et al., 2026 ) . Best results are highlighted in bold .
Table 3: Ablation studies on MoGAP. We study (a) the effectiveness of its key components through ablations and (b) the generalization of MoGAP across WAM architectures. Results report average success rates (%) on GR-1 Tabletop. Best results within each comparison are highlighted in bold .
Figure 5: RoboCasa-Kitchen success rate (%) across execution timesteps.
Figure 6: Real-world experiment success rate (%). Results on the OpenArm humanoid real-world benchmark. We report success rates with standard errors for each method across task suites.
Figure 7: Object relocation task example. The target object’s location is moved during execution. In set II, the object is relocated farther than in set I. MoGAP successfully adapts its trajectory.
Figure 8: Qualitative analysis. The columns show, from left to right, current and future ground-truth observations, latent motion scores, action-to-vision attention maps for the baseline and MoGAP, and their generated future observations.
Table 4: Quantitative analysis. We conduct additional analyses of the latent motion signal, and the trained policy models on GR-1 Tabletop.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Method
24 Tasks Average
Cosmos3 (22K)
55.8
MoGAP (16K)
62.3
Appendix
Table 5: Compute-matched comparison
Figure 9: Task examples of (a) RoboCasa GR-1 Tabletop, (b) RoboCasa-Kitchen, and (c) the multi-view observation input configuration.
Figure 10: In-domain task examples.
Figure 11: Unseen objects task examples.
Figure 12: Unseen environments task examples.
Figure 13: Object relocation task examples (left: Set I; right: Set II)
Figure 14: Visualization of the OpenArm dataset distribution.
World Action Models (WAMs) learn representations for robot control by coupling future visual prediction with action generation. Successful manipulation depends primarily on how the robot and task-relevant objects evolve, whereas future-video prediction must also reconstruct the appearance and background of the entire scene. We therefore propose Trace-WAM, which augments video-action learning with future 3D motion prediction as an auxiliary objective. This objective supervises temporal point correspondence together with image-plane displacement and depth change, directly describing the interaction dynamics and spatial changes involved in manipulation. To this end, we adopt a three-branch Mixture-of-Transformer (Tri-MoT) architecture to jointly learn video prediction, motion prediction, and action generation. The motion expert predicts future trajectories from current visual features, allowing motion supervision to shape the representations used for action generation. Because the action expert does not consume predicted trajectories, inference retains the current-observation policy path: a single visual forward pass followed by action denoising. Experiments on RoboCasa-GR1, LIBERO, and LIBERO-Plus demonstrate strong in-distribution performance and out-of-distribution generalization, with particularly clear gains over Fast-WAM under limited demonstrations and distribution shifts.
Xu Hu, Changjie Wu, Linzhe Shi +8
The Hong Kong Polytechnic University · Amap, Alibaba Group · Tsinghua University +2
World-Action Models (WAMs) augment robot policies with future visual prediction, but it remains unclear what the visual modality should learn for control. While photorealistic future prediction provides dense supervision, it also incurs substantial computation and can allocate capacity to texture, illumination, and background variations that are only weakly related to action selection. Recent efficient WAM variants suggest that the main benefit of the video branch may not lie in the rendered future itself, but in the control-relevant visual representations induced during training. In this work, we revisit future video prediction from a dynamic-centric perspective and ask whether an existing RGB-based WAM can be redirected from appearance-dominated reconstruction toward interaction-induced visual dynamics without introducing additional modality-specific predictions or online inputs at deployment. We propose DC-WAM, a dynamic-centric WAM framework that redistributes supervision and computation in the RGB video branch. At the supervision level, DC-WAM combines temporal-difference flow matching with trajectory-guided weighting, emphasizing dense temporal changes and localized regions where the gripper, manipulated objects, and contact areas move. At the reasoning level, DynaRoute predicts token-wise dynamic relevance and converts it into an attention bias, guiding the model toward control-relevant future tokens. Experiments in simulation and on real-world manipulation tasks show that DC-WAM consistently improves policy performance, especially under out-of-distribution perturbations in lighting, object appearance, and background texture.
Haoyuan Ji, Lingxiang Fan, Shang Su +4
Tsinghua University · Dense-AI · University of Michigan
World action models (WAMs) have emerged as an important class of generalist robot policies, aiming to transfer video prediction priors to action learning. However, we find that existing WAMs frequently produce incorrect motion predictions in open-ended environment, leading to erroneous actions. We attribute this limitation to two factors: (1) base video models are not optimized for manipulation, and (2) naively incorporating action components into video models can substantially degrade their generalization capabilities. We introduce Video Prediction Policy 2 (VPP2), a WAM that enables strong zero-shot generalization in both video prediction and action generation. First, we curate a large-scale, diverse dataset of manipulation videos to continue pretraining the base video foundation model. We annotate video clips with detailed captions and perform event-level video pretraining to promote generalization across open-ended manipulation tasks. Second, we post-train and distill the video model into a single-step visual planner with fixed prediction horizon. Finally, we introduce action module via a mixture-of-transformers (MoT) architecture to learn implicit inverse dynamics model. Experiments demonstrate three key results: (1) VPP2-14B outperforms Cosmos3-64B by 11.0% points in video prediction instruction-following success rate on open-ended tasks; (2) VPP2 surpasses the strongest baseline by 18.5% points in success rate on real-world zero-shot ALOHA manipulation tasks; and (3) following benchmark-specific post-training, VPP2 achieves the highest success rates among evaluated methods on the challenging LIBERO-Pro, LIBERO-OOD, and RoboDojo benchmarks.