Daxiao Robotics and NTU's S-Lab published Mundo, a world model for turning a few images or a short video into a 3D environment that can be explored and acted on. From one to twelve input images the system first builds a 3D proxy that fixes the spatial relations between views, then generates new viewpoints along a chosen camera path; regions that were never photographed are inferred, and more reference views reduce the uncertainty of that fill-in. Larger spaces are handled by generating dozens of context images in segments and stitching them through overlapping frames, which the team says supports continuous paths of tens of seconds to minutes. Objects are described by a scene program holding name, size, pose, mass and friction; scripted forces are run through a physics engine to produce trajectories, contact points and collision forces, and the results are rendered back into the scene with the reference images. The same environment is used for robot navigation, where the 3D proxy supplies object categories, positions, obstacles and traversable regions to a path planner, and for manipulation tasks such as placing an apple in a bowl or stacking blocks, with first-person, third-person and head and wrist camera views generated alongside the task trajectory. A time axis lets a moment be paused while the camera moves, for a bullet-time view of contact and occlusion. The team links a project page and a Hugging Face blog post; no independent evaluation of physical fidelity is reported.
Our takeThe design worth noting is the division of labour: geometry and state in a 3D proxy and scene program, motion in a physics engine, appearance in a video model. That is more inspectable than an end-to-end video generator, because contact forces and object poses are numbers you can check instead of pixels you have to trust. Whether the computed physics matches a real room is precisely what a demo reel cannot show.