LingBot-World: an open-source real-time world simulator
robbyant just open-sourced LingBot-World, a real-time interactive world model built on Wan2.2. It generates controllable 480P and 720P environments from text or images at 16 FPS with sub-second latency, while retaining consistent scene memory over minutes. Why it matters: it gives developers a practical open alternative for interactive worlds and robot learning.
LingBot-World is an open-source, real-time world simulator that transforms a single static image or text prompt into an interactive 3D-like scene you can explore on the fly.
Developed by the robbyant team and built on top of Wan2.2, the project provides a publicly accessible foundation for interactive environments, dynamic game prototyping, and robot learning.
What it can do
The simulator generates interactive scenes at 16 frames per second with sub-second latency, delivering smooth real-time control at 480P and 720P resolutions.
Once an initial scene is generated, users can navigate with keyboard controls and edit the environment mid-session through text prompts, such as switching weather, swapping backgrounds, or introducing new characters. It handles multiple visual styles seamlessly, spanning photorealistic landscapes, stylized animation, and fantasy settings like underwater dives or aerial dragon flights.
Crucially, LingBot-World addresses memory drift with minute-level contextual consistency. In many video generation setups, walking away from a landmark causes the model to forget what was there. LingBot-World maintains spatial permanence: walk around a corner and return, and the structures stay intact.
Hardware specs and how to run it
The system supports flexible control inputs. It offers dedicated weights for camera pose guidance via trajectory matrices, as well as an action-driven variant (Base Act) and an optimized Fast checkpoint.
The codebase and model checkpoints are available on GitHub, Hugging Face, and ModelScope. It requires PyTorch 2.4 or higher and flash-attn.
Local deployment requires serious hardware. The official reference scripts rely on multi-GPU setups running eight cards alongside DiT and T5 FSDP and Ulysses parallelism. Solo developers without cluster access will find local inference heavy, though an interactive web demo hosted by Reactor is available online for immediate testing.
Why it matters
While proprietary systems like Google's Project Genie keep interactive generation behind closed doors, LingBot-World puts functional weights and inference scripts into open hands.
Local compute demands remain steep, but having an open-source world model capable of real-time control and long-term scene persistence gives robotics researchers and indie developers a concrete playground to build on.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com