Daxiao Robotics and Nanyang Technological University Jointly Release World Model Mundo

Recently, Daxiao Robotics, in collaboration with the S-Lab at Nanyang Technological University, released Mundo, a world model built for specific environments that can be explored and interacted with. Mundo supports scene reconstruction, free-viewpoint exploration, and spatiotemporal simulation capabilities similar to those of Atlas. Compared with Atlas, Mundo can incorporate physical properties such as object mass and friction into scene programs, and calculate motion, contact, and collisions through a physics engine. At the same time, Mundo supports calling agents to edit objects, properties, and actions in the scene through code, turning 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. Mundo combines three-dimensional scene representation, physical simulation, and video generation, and works with Astra to create a new spatial understanding capability: it first passes through imagery scenes and objects, and can make the scene editable through code, then defines and edits object properties and actions through code. The physics engine calculates motion and collisions, and the video model combines reference imagery to render the results, forming a complete pipeline from scene reconstruction and code editing to interactive simulation. The scene can not only be seen, but also 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 scene adaptation and policy optimization. Building a dedicated world model from a small number of images, supporting long-path generation across spaces. Mundo expands a small number of images or videos into a three-dimensional space that can be continuously explored; the same set of reference images can support multiple camera paths. For larger scenes, it can also dynamically select context and continuously generate exploration footage ranging from tens of seconds to minute-level duration. Specifically, Mundo starts from contextual imagery of real locations, first builds a three-dimensional proxy, determines the spatial relationships between different views, and then generates 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. The same set of reference images can serve multiple paths, allowing the observation position and angle to change freely. During modeling, a three-dimensional detector provides initial positions and extents for the main objects, and a vision-language model agent further adjusts the proxy scene, through rendering, comparison, and correction, so that the scene structure corresponds to the input imagery. For example, an office can be explored from beside a desk around to a window, and a residence can be entered from the courtyard into the interior. When the camera moves, the input images constrain the layout, appearance, and details that have already been captured, while the model infers and completes areas that have not been captured. Adding reference views provides more real evidence for the scene and reduces the uncertainty of completion. This capability can also start from a single generated image or painting, extending its structure and visual style to new viewpoints. For real locations, more observations help the model approach the actual site; for creative scenes, a single image provides a starting point for building an explorable space. For larger environments, Mundo uses dozens of contextual images to support long-path generation ranging from tens of seconds to minute-level duration. It generates footage segment by segment along the same three-dimensional proxy, selects relevant reference views for each segment, and connects preceding and following segments through overlapping frames, so that local appearance and global space share a common basis. Mundo organizes the individual captured viewpoints into a space that can be continuously entered and explored. For robot scene adaptation, this means that on-site imagery can become an entry point for building task environments, providing a spatial foundation for subsequent navigation and interaction training. Objects have physical properties, and 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, or an object falling to the ground all require corresponding position changes and contact relationships. Mundo can edit the world through code, describe an object's name, size, and pose through scene programs, and configure physical properties such as mass and friction for it. Scripts define actions and applied forces, and the physics engine calculates the object's response under gravity and contact constraints, obtaining motion trajectories, contact positions, and collision forces. Mundo then combines environmental reference images to render these changes into the corresponding scene. This gives the environment clear operational entry points. Users can drag a trash can, push a monitor, pull a chair, or open a drawer, and can also change object materials and gravity conditions to observe responses under different settings. States such as speed, contact points, and impact forces during the physical process can also be displayed synchronously with the generated imagery. In this process, the three-dimensional proxy and scene program are responsible for recording "where objects are and what changes have occurred," the physics engine is responsible for calculating motion, and the video model is responsible for presenting visual results consistent with the environment. Geometry, physics, and appearance each assume clear responsibilities and together form an interactive environment. Making every interaction have a traceable state and a computable process is the key for a scene to move from being viewable to being usable. Mundo introduces physical interaction into reconstructed environments, providing a concrete basis for robots to practice actions and analyze results. Connecting navigation, manipulation, and four-dimensional simulation in the same environment to build a robot training ground. In the same customized scene, Mundo further supports robot movement and manipulation, connecting environment reconstruction to robot testing and training. In terms of navigation, the three-dimensional proxy provides information such as object categories, positions, obstacles, and passable areas, and the path planner uses this to find routes. Mundo can generate first-person and third-person views of different robots traveling along routes, linking the spatial layout with what the robots observe along the way. On the operation side, Mundo establishes geometric and physical properties for the objects involved in interactions, allowing robots of different forms to attempt the same task, such as placing an apple into a bowl, putting a can onto a tray, or stacking blocks. The system generates external views as well as views from the robot's head and wrist cameras, while preserving task trajectories and physical interaction states, providing support for training data generation and simulation testing. Visual observations and interaction states correspond to each other, enabling researchers to analyze the execution process by combining what the robot "sees" with what it "does," and to compare and adjust strategies before entering real scenarios. Mundo also extends spatial exploration into the temporal dimension. For dynamic events recorded by multiple cameras, the process can be presented from new viewpoints; for simulated events constructed from images, motion is computed through gravity and contact dynamics. By pausing time at a selected moment and then moving the camera, the same instant can be observed from multiple angles, forming a "bullet time" effect that helps examine action details and occlusion relationships. From navigation to manipulation, and then to four-dimensional spatiotemporal observation, Mundo provides robots with an environment that can be repeatedly rehearsed and inspected from multiple angles. In a hotel application, the team has used it to reconstruct floors and laundry areas, providing spatial and semantic information for navigation, and to simulate interactions such as opening a washing machine door, supporting robots in understanding tasks and conducting training. The industrial value of a world model must be demonstrated by its contribution to real tasks. In the world model technology system of Daxiao Robotics with Kairos at its core, Mundo supplements specific environment modeling and simulation training capabilities, connecting on-site data collection, data generation, and strategy optimization, and serving robots as they enter actual work scenarios.