Horizon Robotics Releases LoFT Native Video Geometry Foundation Model, Multiple Metrics Reach SoTA
IT Home, November 19 news: On November 18, the Horizon Robotics Research Lab released the native video geometry foundation model LoFT (Looped Foundation Transformer for Video Geometry). According to the introduction, the model can jointly predict camera poses, intrinsics, and per-frame depth from RGB video, recover the three-dimensional structure of a scene, and provide a geometric foundation for robot spatial perception, environment understanding, and embodied learning.
According to the official statement, LoFT is natively designed for continuous video modeling and adopts a dual-loop Transformer architecture, combining cross-frame visual information aggregation with iterative optimization of camera states to mine temporal correlations and multi-view geometric information in video.

IT Home learned from officially published information that the first loop aggregates cross-frame information through local and global temporal attention, gradually expanding the temporal receptive range and leveraging the continuity of adjacent frames and geometric correlations across views; the second loop explicitly updates camera poses and intrinsics, and feeds the camera states back into the next round of feature computation.
The model combines dual-loop features to predict per-frame depth and its confidence, recovering the visible three-dimensional structure of the scene.

According to the official statement, to support large-scale training, the team built a unified 3D reconstruction data processing and training platform, with approximately 2.97 million valid training sequences and 962 million image records, covering diverse scenarios such as simulated and real, dynamic and static, and providing a data foundation through unified data cleaning, quality filtering, and sampling strategies.




In terms of evaluation, the research team compared LoFT with representative models such as VGGT-Ω and DA3-Giant across three evaluation systems: SpatialBench, ClassicBench, and VRBench. Among them, SpatialBench covers 12 datasets and 98 continuous video scenes; ClassicBench integrates 7 classic video reconstruction benchmarks; and VRBench covers robotics, autonomous driving, and general video scenarios. According to officially published information, LoFT achieves SoTA on multiple depth estimation metrics.


On the application side, the official statement says LoFT can use temporal and multi-view geometric information in continuous video to recover the three-dimensional structure of scenes and camera motion trajectories, covering scenarios such as robot first-person video, dynamic streets, indoor manipulation, and open-environment exploration.
Advertising statement: External redirect links contained in this article (including but not limited to hyperlinks, QR codes, passcodes, etc.) are used to convey more information and save selection time. The results are for reference only. All articles on IT Home that contain external links include this statement.
Related Articles
- The Forza Horizon 6 roadmap for late 2026-2027 is announced, including two DLCs and content updates
- Horizon Robotics founder Yu Kai criticizes inflated computing power claims in smart driving chip PPTs: 英伟达 labels it at twice the dense computing power, while the industry has already inflated it to 6-7 times
- Starting at 79,800 yuan: FAW-Volkswagen Jetta M6 launched, with Horizon Robotics Journey 6M assisted driving
- The PS5 version of Forza Horizon 6 is officially announced for release on January 27, 2027, Beijing time, starting at HK$498
- Forza Horizon 6 developer Playground Games admits insufficient communication with players and will consider adjustments later
- Sources say the PS5 version of Forza Horizon 6 will be released on January 26, 2027, with a physical disc version