Daxia Robotics and Nanyang Technological University Release Mundo, a World Model for Free Exploration and Physical Interaction

This newspaper report (Reporter: Jia Li) recently saw Shanghai Daxiao Unlimited Robotics Co., Ltd. (hereinafter referred to as "Daxiao Robotics") jointly with the S-Lab Laboratory of Nanyang Technological University release Mundo ("World"), a world model designed for specific environments that is explorable and interactive. The model supports scene reconstruction, free-viewpoint exploration, and spatiotemporal simulation, incorporating physical properties such as mass and friction of objects into scene programs, and calculating motion, contact, and collisions via a physics engine. Mundo combines 3D scene representation, physics simulation, and video generation to transform a small number of images into a 3D world that is freely explorable, programmatically interactive, and suitable for robot navigation and operation training, providing environmental modeling, training data generation, and simulation support for robots entering specific commercial scenarios.

Supports long-path generation across spaces

It is understood that Mundo can expand a small number of images or videos into a 3D space that is sustainable for exploration. A single set of reference images can support multiple camera paths; for larger scenes, it can dynamically select context and continuously generate exploration footage lasting from tens of seconds to minutes. Specifically, starting from contextual imagery of real locations, Mundo first establishes a 3D proxy, determines spatial relationships between different views, and then generates new viewpoints along specified camera paths. Inputting 1 to 12 images enables exploration around furniture and spatial structures within a room.

During modeling, the 3D detector provides initial positions and ranges for major objects, while visual language model agents further adjust the proxy scene. Through rendering, comparison, and correction, the scene structure is made to correspond with the input imagery.

For example, in an office, the viewpoint can move from beside a desk to a window; in a residence, it can move from a courtyard into the interior. As the camera moves, input images constrain the layout, appearance, and details already captured, while the model infers and completes unseen areas. Adding reference views provides more real-world basis for the scene, reducing uncertainty in completion. This capability can also start from a single generated image or painting, extending its structure and visual style to new viewpoints.

For larger environments, Mundo uses dozens of contextual images to support long-path generation lasting from tens of seconds to minutes. It generates footage segment by segment along the same 3D proxy, selects relevant reference views for each segment, and connects preceding and succeeding segments via overlapping frames, ensuring that local appearance and global space share a common basis. For robot scenario adaptation, this means on-site imagery can serve as an entry point for constructing task environments, providing spatial foundations for subsequent navigation and interaction training.

Push-pull collisions have computable responses

For robots to practice within a scene, the environment must be able to respond to actions. Mundo allows world editing via code, describes object names, dimensions, and poses through scene programs, and configures physical properties such as mass and friction for them; scripts define actions and applied forces, while the physics engine calculates responses under gravity and contact constraints, yielding motion trajectories, contact positions, and collision forces. These are then rendered into corresponding scenes combined with environmental reference images. This model can provide computable responses to interactions such as pushing, pulling, and collisions.

Within the same customized scene, Mundo further supports robot movement and operation, linking environment reconstruction with robot testing and training. For navigation, the 3D proxy provides information on object categories, positions, obstacles, and traversable areas; a path planner uses this to find routes, and Mundo can generate first-person and third-person footage of different robots traveling along these routes. For operation, Mundo constructs geometric representations for objects involved in interaction and configures their physical properties, allowing robots of different forms to attempt the same task—for example, placing an apple into a bowl, putting a can onto a tray, or stacking building blocks. The system generates footage from external viewpoints as well as robot head and wrist camera perspectives, while retaining task trajectories and physical interaction states.

From navigation to operation, and further to four-dimensional spatiotemporal observation, Mundo provides robots with an environment that can be repeatedly practiced in and checked from multiple angles. In hotel applications, the team has already used it to reconstruct floors and laundry areas, providing spatial and semantic information for navigation, and simulating interactions such as opening a washing machine door, supporting robot understanding of tasks and training.

In Daxiao Robotics' world model technology system centered on Kairos ("Enlightenment" World Model), Mundo supplements capabilities in specific environment modeling and simulation training, connecting on-site data collection, data generation, and strategy optimization to serve robots entering actual work scenarios.

(Editor: Wang Hao)