Mundo turns a few photos into an explorable, editable world for robot simulation
ccloy · x · 2026-10-23
A new project led by onelineluo (advised by ccloy) introduces Mundo, a customized world model that builds an interactive 3D environment from just 1-12 images of a place:
- Spatial exploration: generates new views along camera paths, from sparse views up to long trajectories through large scenes with dozens of context images
- Physically grounded interaction: models object geometry and physical properties; simulates scripted actions and applied forces with rendered motion and collisions
- Dynamic simulation: simulates agents and objects interacting over time, rendered from external and first-person viewpoints
- Robotics: supports Real-to-Sim workflows for testing and training robot navigation and manipulation in the same environment
The key is an explicit, editable world state: code specifies actions, physics computes consequences, and a video model renders outcomes using input images as visual context. Code coming soon.
More from Embodied
- Ropedia releases DONIC: one policy controls two dexterous robot hands, 86.79% success on unseen trajectories — liuziwei7 · 2026-10-23
- Ropedia A0 runs robot policies as pure code, hitting 70.24% on RoboDojo vs 46.85% baseline — liuziwei7 · 2026-10-23
- Agile Robots and NVIDIA unveil Agile ONE SN humanoid, a European answer to Unitree G1 — CyberRobooo · 2026-10-23
- A pneumatic muscle-driven artificial finger demo worth watching — repligate · 2026-10-23
- Tesla adds Las Vegas, Phoenix, Jacksonville, San Antonio to Robotaxi map — VectorWireAI · 2026-10-23
- PDW integrates Booz Allen's Shield AI-powered autonomy into new drone platform — BrettKrieger12 · 2026-10-23