X-WAM: Bridging Action and 4D Synthesis

Current unified world models struggle to bridge the gap between efficient real-time robotic action and high-fidelity 4D world synthesis. Existing approaches often remain confined to 2D pixel spaces, compromising either action execution speed or the quality of spatial understanding.

Unifying Action and 4D Synthesis with X-WAM

The X-WAM unified 4D world model emerges as a singular framework to address this dichotomy, integrating real-time robotic action execution with high-fidelity 4D world synthesis, encompassing both video and 3D reconstruction. This novel architecture leverages the potent visual priors of pretrained video diffusion models to project future world states. A key innovation is its efficient spatial information extraction via a lightweight structural adaptation: replicating the final blocks of a pretrained Diffusion Transformer to a dedicated depth prediction branch. This design allows for the reconstruction of future spatial information without the prohibitive computational overhead of full 3D modeling.

Asynchronous Denoising for Enhanced Efficiency and Quality

To further optimize performance, X-WAM introduces Asynchronous Noise Sampling (ANS). This technique tackles the critical trade-off between generation quality and action decoding efficiency. During inference, ANS employs a specialized asynchronous denoising schedule. This allows for rapid action decoding using fewer steps, facilitating real-time execution. Concurrently, the full denoising sequence is reserved for generating high-fidelity video. Crucially, ANS aligns the training and inference distributions by sampling from the joint distribution of timesteps, rather than decoupling them entirely during training, ensuring robust performance across both tasks.

Empirical Validation and Strategic Implications

Pretrained on an extensive dataset of over 5,800 hours of robotic data, X-WAM demonstrates significant advancements. It achieves impressive average success rates of 79.2% on the RoboCasa benchmark and 90.7% on RoboTwin 2.0. Beyond task completion, X-WAM delivers superior 4D reconstruction and generation quality, outperforming existing methods in both visual and geometric metrics. This breakthrough positions the X-WAM unified 4D world model as a foundational technology for next-generation embodied AI systems, enabling more capable and visually rich robotic agents.