Mundo

Mundo is a customized robot world model released on October 30, 2026 by ACE Robotics (大晓机器人) together with NTU S-Lab (Chen Change Loy's group). From a few photos or a short video of a real place, it rebuilds an explorable, editable 3D world where objects have physical properties and robots can navigate and manipulate.

Under the hood it combines an explicit 3D proxy, a code-based scene program, a physics engine (MuJoCo) and a video diffusion model that renders what happens using the original photos as visual context.

What it can do

• Free-view exploration: 1–12 input images are enough to fly a camera around a room; dozens of images support tens-of-seconds to minute-long paths through larger spaces.
• Physics: objects get mass, friction and other properties in a scene program; pushes, pulls, collisions and gravity are computed by the physics engine and then rendered.
• Agent-editable: a VLM agent can edit objects, properties and actions through code.
• Robotics: supports navigation (route planning on the 3D proxy) and manipulation (e.g. put an apple in a bowl) with first-/third-person and wrist-camera views, keeping trajectories and contact states for Real-to-Sim data generation.
• 4D "bullet time": pause a simulated moment and orbit the camera.

Context & limits

Mundo is part of ACE Robotics' Kairos (开悟) world-model stack. The authors note it is still slow (scene reconstruction 10–20 minutes; rendering ~1.5 fps on eight A800 GPUs), small objects can float, and there is no long-horizon memory yet. It launched the same day as Qiyuan's PrimeSim.

Where to start

1. Project page with videos: mundo-world-model.github.io.
2. Technical write-up: Hugging Face blog — Mundo.
3. Useful for robotics teams building Real-to-Sim pipelines (e.g. reconstruct a hotel floor, then train navigation/manipulation), and for researchers in 3D reconstruction + video generation.
4. Cite as luo2026mundo ("Mundo: Customized World Model").