THE ESSENTIALS
  • LoFT targets camera geometry and visible scene reconstruction from continuous RGB video.
  • Developer-reported evaluations and possible uses do not establish independent validation, commercial deployment or licensing.

Geometry from continuous video

Horizon Robotics' robotics laboratory released LoFT, short for Looped Foundation Transformer for Video Geometry, on November 18, IT Home reported on November 19. The model jointly predicts camera poses, camera intrinsics and frame-by-frame depth from RGB video to reconstruct visible three-dimensional scene structure.

According to the information described by the outlet, LoFT uses a dual-loop Transformer. One loop aggregates information across frames through local and global temporal attention. The other updates camera parameters and feeds that state back into subsequent feature computation.

Research claims and limits

The team describes a training platform covering approximately 2.97 million sequences and 962 million image frames, spanning simulated and real scenes. Its reported evaluations compare LoFT with other reconstruction models across three benchmark suites. Those results are attributed to the developers; this report has not reproduced the experiments.

The stated applications include robot first-person video, indoor manipulation and dynamic streets. These are proposed perception and reconstruction uses, rather than evidence that the model has been deployed in commercial robots or validated as an autonomous-driving system. The report does not establish download availability or licensing terms.

Sources & context

Go to the original material. Company claims remain attributed to their sources.

01
地平线发布 LoFT 原生视频几何基础模型,多项指标达到 SoTA - IT之家 ↗Retrieved and checked by the China AI Wire agent; evidence ledger retained internally.
Last updated November 19, 2026.Spotted an issue? Let us know ↗