Recently, Daxiao Robotics, together with Nanyang Technological University's S-Lab, released Mundo, a world model built for specific environments that can be explored and interacted with.
Building on Atlas scene reconstruction, free-viewpoint exploration, and spatiotemporal simulation, Mundo additionally incorporates physical properties such as object mass and friction into the scene program, computing motion and collisions through a physics engine; it also supports calling agents to define object properties and actions in code, and, combined with video model rendering, turns a small amount of imagery into a three-dimensional world that can be freely explored, programmatically interacted with, and used for robot navigation and manipulation training.
The scene can both be seen and respond to actions and record changes, thereby becoming an environment for robot training and testing.
This capability provides support for robots entering specific commercial scenarios. In Daxiao Robotics' world model technology system centered on Kairos, Mundo supplements environment modeling, training data generation, and simulation training capabilities, helping robots complete scenario adaptation and policy optimization.
Project:
https://mundo-world-model.github.io/
Blog:
https://huggingface.co/blog/Luo-Yihang/mundo-world-model
Building a dedicated world model from a small number of images
Supports cross-space long-path generation
Mundo expands a small number of images or videos into a continuously explorable three-dimensional space; the same set of reference images can support multiple camera paths; for larger scenes, it can also dynamically select context to continuously generate tens of seconds to minutes of exploration footage.
Specifically, Mundo starts from contextual imagery of real places, first building a three-dimensional proxy to determine the spatial relationships between different views, and then generating new viewpoints along a specified camera path.
In the scenes demonstrated this time, inputting 1 to 12 images is enough to explore around the furniture and spatial structure in a room.
For example, an office can be explored from beside a desk around to a window, and a residence can be explored from the courtyard into the interior. As the camera moves, the input images constrain the layout, appearance, and details that have already been captured, while the model infers and completes the areas that have not been captured. Adding reference views provides more real evidence for the scene and reduces the uncertainty of completion.
This capability is not limited to real places: a single painting or generated image can have its structure and style extended into an explorable three-dimensional space. For larger environments, Mundo uses dozens of context images to generate in segments, connecting successive segments with overlapping frames to support continuous long-path exploration of tens of seconds to minutes.
Mundo organizes the individual captured viewpoints into a space that can be continuously entered and explored. For robot scenario adaptation, this means that on-site imagery can become the entry point for building a task environment, providing a spatial foundation for subsequent navigation and interaction training.
Objects have physical properties
Pushing, pulling, and collisions all have computable responses
For a robot to practice in a scene, the environment must be able to respond to actions. A drawer being pulled open, a chair being pushed, an object falling to the ground—all require corresponding position changes and contact relationships.
Mundo can edit the world through code, describing an object's name, dimensions, and pose through a scene program, and configuring physical properties such as mass and friction for it. Scripts define actions and applied forces, and the physics engine computes the object's response under gravity and contact constraints, yielding motion trajectories, contact positions, and collision forces. Mundo then combines these with environment reference images to render these changes into the corresponding scene.
This gives the environment a clear entry point for manipulation.