Robbyant LingBot-World is an open-source framework for interactive world modeling from Robbyant, an embodied intelligence group. Its base model, LingBot-World-Base, produces high-fidelity, controllable and logically consistent simulations. Instead of passive video synthesis, it learns physics and causality from large-scale gaming environments through a Scalable Data Engine, so users can interact with the generated worlds. Fine-grained, action-conditioned generation responds to commands and renders physically plausible dynamic scenes. Code, a tech report and model weights are published on Hugging Face and ModelScope.
The release is aimed at AI researchers, robotics and embodied intelligence teams, game developers and simulation engineers who study or build world models. Because the model treats game engines as infinite data generators and aims to generalize from synthetic data to real scenarios, it is relevant to teams that need training environments for agents. Developers can inspect and adapt the open code rather than rely on a closed service.
Capabilities described on the project page include consistency over minute-long trajectories, with object permanence and narrative logic, and dynamic off-screen memory, where an agent such as a cat keeps acting while out of view. Grounded physical constraints stop agents from clipping through obstacles. LingBot-World-Fast offers low-latency inference for real-time closed-loop control, turning a video into a playable simulator. A demo gallery lets visitors pick a world setting and an event, such as fireworks, a storm or ice, to see the future generated, and an Action Agent plans and executes actions inside the generated world.
