LingBot-Map is an open-source 3D foundation model from Ant Group’s Robbyant team, released on GitHub where it has gathered more than 18,100 stars. It performs streaming 3D reconstruction: feed it a video or a folder of images and it reconstructs the scene’s geometry and camera trajectory in a single feed-forward pass, at around 20 frames per second over sequences longer than 10,000 frames. The work is an ECCV 2026 Best Paper Award candidate, and the repository is actively updated in 2026.
The model’s Geometric Context Transformer unifies coordinate grounding, dense geometric cues, and long-range drift correction inside one streaming framework, using anchor context, a pose-reference window, and trajectory memory. Paged KV-cache attention keeps inference stable and memory-bounded on very long sequences, with keyframe and windowed modes for limited GPU memory and an SDPA fallback when FlashInfer is unavailable. Checkpoints are published on Hugging Face and ModelScope, an interactive viser-based demo renders point clouds in the browser, and an offline rendering pipeline handles long videos such as a 25,000-frame indoor walkthrough. Evaluation code covers standard benchmarks including KITTI, TUM-D, ETH3D, and Tanks and Temples.
LingBot-Map is free and open source under the Apache-2.0 license, including the released model checkpoints.
LingBot-Map is for robotics, AR/VR, and computer-vision researchers and engineers who need fast, scalable 3D reconstruction from ordinary video without iterative optimization pipelines. Its streaming design also makes it a practical building block for embodied AI systems that must build a live 3D map of their surroundings while moving through them.