Horizon Robotics' Robotics Laboratory has announced the release of LoFT, a native video geometry foundation model capable of jointly predicting camera pose, intrinsics, and per-frame depth directly from RGB video. The model reconstructs the 3D structure of a scene, providing a geometric foundation for robot spatial perception. LoFT employs a dual-loop Transformer architecture: the first loop aggregates cross-frame information, while the second updates camera states and feeds them back into feature computation. By combining features from both loops, the model predicts per-frame depth and confidence. The team built a dedicated training platform and completed training on nearly 2.97 million valid sequences—amounting to 962 million frames—covering images from a wide range of scene types. In comparative testing across three evaluation systems, LoFT achieves state-of-the-art performance on multiple depth estimation metrics and adapts to diverse scenarios, including robot first-person video and dynamic street scenes.
