robbyant/lingbot-map

LingBot-Map is an AI model that takes a stream of images, like frames from a video of someone walking through a building, and reconstructs a 3D map of that environment in real time. Think of it as software that watches footage and builds a navigable 3D model as it goes, rather than needing all the footage upfront. The problem it solves: most existing methods for turning images into 3D scenes either require processing everything offline (slow, can't keep up with live data) or lose accuracy over long sequences as small position errors accumulate. LingBot-Map processes frames one by one as they arrive and uses a built-in memory system to correct drift, keeping the 3D reconstruction accurate even over very long videos, the README mentions sequences of over 25,000 frames (about 13 minutes of indoor footage). How it works: the model uses a special architecture called a Geometric Context Transformer (GCT) that tracks where the camera is in space while simultaneously building the 3D map. A paged KV cache, a technique borrowed from how AI language models manage memory efficiently, lets it run at roughly 20 frames per second without running out of GPU memory. You provide a folder of images or video frames, and it produces an interactive 3D viewer you access in a browser. You would use it in robotics, autonomous navigation, AR/VR scene capture, or any research context where you need to reconstruct environments from video in real time. It's a research-grade Python project. The full README is longer than what was provided.