research

Horizon Robotics unveils LoFT, a 514.5M-parameter video geometry model; paper, code and weights still pending

TL;DR

Horizon Robotics announced LoFT on November 18, a video-geometry foundation model with approximately 514.5 million parameters that predicts camera pose, camera intrinsics and per-frame depth from RGB video. The project page still lists the paper, code and Hugging Face resources as Coming soon.

3 min read
0

Horizon Robotics announced LoFT on November 18, a video-geometry foundation model that predicts camera pose, camera intrinsics and per-frame depth from RGB video. The project page lists approximately 514.5 million parameters. The paper, code and Hugging Face resources are all still marked "Coming soon," so there is no open-source package to download yet. Pandaily reported the announcement.

What LoFT is

LoFT stands for Looped Foundation Transformer for Video Geometry. According to Horizon Robotics, it uses continuity across neighboring frames and information from different viewpoints to reconstruct a scene's visible three-dimensional structure. It is a geometry-perception model. It is not a video-generation service or a ready-made robot-control policy. Horizon Robotics presents it as a route toward richer spatial perception for robots.

Architecture

The team describes a two-loop Transformer design.

  • Feature loop: Visual features are aggregated through progressively wider local temporal windows, expanding from three frames to five, nine and 17. These local steps alternate with attention across the full clip, so distant views of the same location can contribute information.
  • Camera loop: The model iteratively updates camera state, including pose and field of view, then feeds that state back into the next feature update.
  • Output head: A dense prediction head produces depth and confidence for each frame.

Each loop reuses its own weights across iterations. The project says the same model can therefore run at different refinement depths.

Training setup

The project page reports training on 2.97 million sequences containing 961.8 million images, drawn from a mixture of 80 public datasets. These are the team's reported figures. They have not been independently audited.

Evaluation

LoFT is evaluated on SpatialBench, ClassicBench and VRBench. Specific scores are not given in the announcement text. The framing differs between the company's two descriptions:

  • Horizon Robotics' Chinese announcement reports leading results on selected depth-estimation metrics.
  • The English project abstract describes overall depth and camera-pose results as competitive.

Performance varies between depth, pose and trajectory metrics, so the published comparisons should not be read as an across-the-board state-of-the-art claim.

Demonstrations and intended uses

The company shows reconstructions of street scenes, indoor activity and robot-related observations. It lists spatial understanding, environment exploration and embodied learning as potential downstream uses. These demonstrations do not establish deployment reliability for an autonomous robot.

Availability

Public descriptions and demonstration assets can be inspected. The project has not yet supplied the paper, runnable code or downloadable weights through its listed resource controls. Pricing, licensing and release terms have not yet been published.

What this means

LoFT is currently an announcement with reported specifications, not a tool researchers can run. The architecture is the most concrete part of the release: reusing weights across loop iterations means refinement depth can be adjusted without changing the model, according to the project. The training figures (80 public datasets, 961.8 million images) show the scale of the reported setup.

The mixed framing of results, leading on selected depth metrics in the Chinese announcement and competitive overall in the English abstract, makes the paper the key document. Until it is published, the benchmark claims cannot be checked against metric-level tables. Independent reproduction depends on the code, weights and license terms that the project has yet to release.

Related Articles

research

ACE Robotics' ACE-CaP-0 lets AI agents write reusable robot code, reporting 92.1% success

ACE Robotics has released ACE-CaP-0, a research recipe in which a general-purpose AI agent writes, tests and refines robot control programs, then freezes them into reusable skills that run without the agent. The work reports 92.1% success. The paper was posted to arXiv on November 12.

research

AWS releases I/Q, I/R and Isabelle Assistant to connect AI agents to the Isabelle proof checker via MCP

AWS has made three tools in its AutoCorrode verification suite publicly available: I/Q, I/R and Isabelle Assistant. They connect AI agents to the Isabelle proof assistant through the Model Context Protocol, so generated proofs are checked by computer. Amazon Science authors demonstrate the setup using Claude Opus 4.6.

research

OpenAI Reports Up to 14% PPA Gain on AMD Flagship GPU IP From Model-Rewritten RTL

OpenAI reported on November 18, 2026 that its models achieved up to a 14% improvement in power, performance and area (PPA) on AMD flagship GPU intellectual property, measured against heavily optimized baselines. Some of the circuit rewrites were included in a design sent for manufacturing.

research

OpenAI models optimized circuits used in Jalapeño silicon and improved RISC-V core designs

OpenAI reports that its models optimized circuits used in Jalapeño silicon and improved RISC-V core designs. The company presents the work as progress in AI-assisted chip design with GPT-6 Astra.

Comments

Loading...