Ant Group's LingBot has officially open-sourced LingBot-Map, a groundbreaking foundational model designed for streaming 3D reconstruction. This innovative model enables real-time camera pose estimation and 3D scene reconstruction using just an ordinary RGB camera, showcasing remarkable robustness in processing long sequences of data. Since its release, LingBot-Map has not only attracted widespread attention on GitHub but has also earned a coveted spot for an oral presentation at ECCV 2026.

Following its open-sourcing, developers worldwide have enthusiastically embraced LingBot-Map for secondary development. Their efforts span optimizing deployment workflows, tailoring the model for consumer-grade hardware, and seamlessly integrating it with cutting-edge devices like smart glasses. Meanwhile, the official team remains committed to continuously refining the model to boost its practical usability.

LingBot-Map introduces a pioneering three-layer Geometric Context Attention (GCA) architecture, which effectively resolves the inherent contradiction between memory and computational storage demands in streaming reconstruction. This architectural innovation has enabled the model to excel across multiple benchmark tests, securing leading positions in key metrics such as pose accuracy, trajectory consistency, and reconstruction quality.

As a core component of Ant Group's LingBot spatial perception system, LingBot-Map works in tandem with LingBot-Vision and LingBot-Depth to form a comprehensive spatial intelligence pipeline. This pipeline spans the entire spectrum from perception to understanding and ultimately to action, empowering robots to transcend mere 'visual perception' of the world and instead 'comprehend' and 'leverage' it in meaningful ways.