Robbyant LingBot-World

Robbyant LingBot-World is an open source world model for embodied AI that generates controllable, physics aware game and physical environments, supports real time interactive control through a fast variant, keeps object permanence over minute long runs, and reconstructs 3D scenes from generated sequences.

LingBot-World is an open source world model from Robbyant, built for interactive simulation rather than passive video output. Its base model (LingBot-World-Base) learns from large volumes of gaming environments through a Scalable Data Engine, so it acquires physics and causality signals that let a generated scene respond to commands instead of drifting on its own. The claim is fine grained, action conditioned generation: the system reacts to user direction and renders dynamic scenes that hold together physically. In practice the tool behaves like a playable simulator. A user picks a world setting and an event, and the model generates what happens next; the demo gallery shows examples such as a dragon, a fountain, a seabed and an ice world with events layered on top. A Fast variant (LingBot-World-Fast) targets low latency inference for real time closed loop control. Contextual memory keeps object permanence and scene structure across minute long trajectories, including agents that keep acting while off screen, and collision rules prevent characters from clipping through solid obstacles. Generated sequences can also be reconstructed into 3D models that rotate and zoom in the browser. An action agent layer is described as planning and executing actions inside the generated world. The audience is research teams, game and simulation studios, robotics and embodied AI groups, and graphics engineers comfortable with model weights and inference hardware. It is not a plug and play creative tool: the materials point to a GitHub repository, Hugging Face and Model Scope collections, and an arXiv tech report, and real time performance through causal distillation currently costs some visual fidelity. Positioning is closest to interactive world models and neural game engines rather than text-to-video generators. Robbyant is explicit about limits: enterprise grade GPUs are required, memory is emergent from the context window so scenes drift over long runs, controls cover basic navigation only, and object manipulation is not yet supported. The roadmap names explicit memory modules, a wider action space and drift elimination as next steps. Teams wanting a finished production pipeline should look elsewhere for now; research and prototyping groups get an unusually transparent open framework.