Dexterous World Models (2512.17907v1)

Abstract: Recent progress in 3D reconstruction has made it easy to create realistic digital twins from everyday environments. However, current digital twins remain largely static and are limited to navigation and view synthesis without embodied interactivity. To bridge this gap, we introduce Dexterous World Model (DWM), a scene-action-conditioned video diffusion framework that models how dexterous human actions induce dynamic changes in static 3D scenes. Given a static 3D scene rendering and an egocentric hand motion sequence, DWM generates temporally coherent videos depicting plausible human-scene interactions. Our approach conditions video generation on (1) static scene renderings following a specified camera trajectory to ensure spatial consistency, and (2) egocentric hand mesh renderings that encode both geometry and motion cues to model action-conditioned dynamics directly. To train DWM, we construct a hybrid interaction video dataset. Synthetic egocentric interactions provide fully aligned supervision for joint locomotion and manipulation learning, while fixed-camera real-world videos contribute diverse and realistic object dynamics. Experiments demonstrate that DWM enables realistic and physically plausible interactions, such as grasping, opening, and moving objects, while maintaining camera and scene consistency. This framework represents a first step toward video diffusion-based interactive digital twins and enables embodied simulation from egocentric actions.

• The paper introduces a novel framework that explicitly models residual dynamics induced by human interactions in static 3D scenes.
• It leverages a video diffusion model conditioned on egocentric camera trajectories and high-fidelity hand renderings, yielding superior PSNR, SSIM, and LPIPS performance.
• The approach supports simulation-based decision-making and shows robust generalization across synthetic and real-world settings for embodied AI.

Dexterous World Models: Action-Conditioned Visual Simulation in 3D Environments

Problem Formulation and Motivation

The "Dexterous World Models" (DWM) framework addresses a central limitation in embodied AI: while digital twins and 3D reconstructions permit realistic rendering of environments, most are static and lack the capability to simulate interactive human manipulations with high fidelity. Existing world models or video diffusion architectures typically entangle scene synthesis with dynamics modeling, treat navigation as the primary driver of change, or rely on high-level (e.g., language) action conditioning. This restricts their utility for tasks demanding fine-grained, causally-grounded anticipation of human-induced environmental changes.

DWM proposes a new formalism for world modeling: explicitly conditioning video diffusion on both a rendered static 3D scene observed along a known egocentric camera trajectory and a temporally precise egocentric hand trajectory (as mesh renderings). This approach directly models the residual dynamics induced by human interactions – such as grasping or manipulation – while maintaining photorealistic and temporally coherent scene consistency.

Figure 1: Overview of the DWM framework, which simulates egocentric visual dynamics from embodied (hand) actions within a static 3D scene by conditioning a video diffusion model on scene renderings and hand trajectories.

Technical Approach

DWM reframes world modeling as sampling from $$p_\theta(\mathbf{V}_{1:F} \mid \mathbf{S}_0, \{\mathcal{C}_{1:F}, \mathcal{H}_{1:F}\})$$, where $\mathbf{S}_0$ is a static 3D scene, $\mathcal{C}_{1:F}$ the camera trajectory, and $\mathcal{H}_{1:F}$ the sequence of dexterous hand states. The model factorizes dynamics as residual changes $\Delta \mathbf{S}_{1:F}$ on top of the static world, ensuring that unaltered regions persist, while only manipulation-driven changes are synthesized.

DWM leverages a video diffusion model, initialized from a strong video inpainting prior, to act as a near-identity operator on static scene renderings (preserving spatial and temporal structure) and conditions the model on high-fidelity, mesh-based egocentric hand renderings (not merely segmentation masks or pose parameters). This enables precise spatial and articulation cues critical for detailed manipulation synthesis. The video diffusion model operates in the VAE latent space, with conditioning latents concatenated to the noisy latent inputs at each denoising step.

A distinguishing design choice is the explicit disentanglement of camera navigation and hand-induced manipulation. In the absence of hand motion input, DWM simulates navigation only; with hand motion, it modulates scene content accordingly, achieving clear navigation-manipulation separation.

Figure 2: Navigation-manipulation disentanglement: Without hand-motion input, DWM simulates navigation-only; with hand conditioning, realistic action-induced scene changes emerge.

Dataset Construction and Training

Crucial to DWM is access to training triplets: (i) static scene video along a camera trajectory, (ii) egocentric hand-mesh video, (iii) ground-truth interaction video. To overcome the impracticality of collecting such data in real-world, dynamic-view scenarios, a hybrid strategy is used:

• Synthetic Egocentric Interactions: Using simulators like TRUMANS, which provide precisely aligned triplets, enabling fine-grained locomotion-plus-manipulation supervision.
• Fixed-Camera Real-World Interactions: Utilizing TASTE-Rob and other sources, allowing the model to absorb rich, real-world physical dynamics under static views by constructing paired static-interaction sequences.
• Dynamic-View Real-World Evaluation: A new data collection protocol using SLAM-enabled Aria Glasses to reconstruct static scenes and generate aligned egocentric renderings for rigorous evaluation of generalization.

  Figure 3: Example paired real-world data with dynamic view: static scene video and ground-truth interaction video, collected via Aria Glasses and custom protocol.

Empirical Evaluation

DWM is evaluated on a comprehensive benchmark spanning synthetic dynamic-view, real-world static-view, and real-world dynamic-view settings. Comparisons are made against strong baselines: SDEdit (using text-only editing), CogVideoX-Fun fine-tuning (video inpainting with text prompts), and InterDyn (hand-mask conditioned ControlNet-based diffusion).

DWM achieves consistently superior quantitative results across PSNR, SSIM, LPIPS, and DreamSim metrics. Importantly, DWM demonstrates robust generalization to completely out-of-distribution real-world scenes and manipulation types, including unseen articulated objects (e.g., opening windows), without scene or action-specific retraining.

Figure 4: Qualitative comparison on synthetic and real-world dynamic scenes: DWM generates physically plausible, action-conditioned visual dynamics and generalizes to unseen real-world scenarios.

Figure 5: On real-world static-camera scenes, DWM maintains scene consistency and produces correct, meaningful interactions, whereas baselines hallucinate or fail to enact intended changes.

Ablation studies underscore:

• The efficacy of hybrid training data—static-camera real-world samples augment generalization even to dynamic-view tasks.
• Hand-mesh rendering for conditioning clearly outperforms both AdaLN parameter-space and hand-mask approaches in perceptual and pixel-space metrics.
• Model initialization from an inpainting prior offers significantly more stable and performant training for residual dynamics than from image-to-video models.

Action Simulation and Decision-Making

DWM is not just a generative model but supports simulation-based reasoning. Given an initial state and multiple hand action candidates, DWM can simulate and rank their consequences via goal-conditioned metrics (VideoCLIP similarity for language goals, LPIPS for image goals), directly supporting action selection for embodied AI.

Figure 6: Action evaluation capability — DWM simulates candidate actions and ranks them by goal alignment, enabling visual reasoning for decision-making.

Robot-Centric Extensions

DWM outputs can be mapped to robot-scene videos by substituting hand meshes with robot arms using, e.g., the Masquerade pipeline, enabling visual data augmentation for robot learning and policy transfer without explicit physics.

Figure 7: Human-scene interaction videos produced by DWM are converted into robot-scene manipulation videos, supporting scalable visual simulation for robotics.

Implications, Limitations, and Future Directions

DWM establishes a formal and architectural template for simulating environment dynamics explicitly induced by articulated, egocentric action in 3D scenes—a fundamental capability for both embodied AI and interactive digital twins. Practical implications span data augmentation for robotics, embodied planning, and offline reasoning about human-object interaction consequences in photorealistic settings.

Limitations include reliance on text prompts for best semantic control, challenges with highly nonrigid or deformable object interactions, and difficulties in collecting large-scale, dynamically-aligned real-world training data. Explicit geometric or contact modeling is absent, limiting physical constraint enforcement. Future research will likely incorporate depth-aware perception, 3D physical priors, richer interaction datasets, and differentiable simulation for policy optimization.

Conclusion

Dexterous World Models introduce an action- and scene-conditioned video diffusion paradigm for egocentric, fine-grained visual simulation of human-driven scene dynamics. Through hybrid data construction, explicit residual modeling, strong quantitative and qualitative performance, and robust generalization to novel real-world scenarios, DWM advances the development of interactive, manipulation-scaled digital twins and simulation-centric embodied AI systems.

Figure 8: Accurate targeting: DWM manipulates correct target objects in response to hand-action conditioning, affirming fidelity of action-to-dynamics mapping.