What is remot?

ReMoT is the code, dataset, and trained model released alongside a research paper accepted as a Highlight at the CVPR 2026 computer vision conference. The paper is about teaching AI vision systems to understand motion in images, specifically by training them on sets of three related images at a time, called triplets, so the model learns to notice what changed between frames rather than just what is present in a single picture. The dataset is made up of question and answer pairs stored in a simple text format called JSONL, where each entry includes a set of images, a written question, and the correct answer. Many of the examples come from robotics footage, asking the model to figure out whether a robot's gripper opened or closed between frames, or whether a camera's viewpoint shifted between shots. Other examples cover more general visual reasoning tasks, like counting objects, detecting objects, or judging their relative position in a scene. Because only the text files are shared publicly and not the original images themselves, the README explains how to trace each entry back to its source image using patterns in the file names. The repository includes scripts for evaluating a trained model against these datasets. One script evaluates a single dataset across multiple graphics cards at once, another runs the full set of benchmark datasets in one go, and a third script gathers all the results into a summary report. The evaluation measures how often the model's answer exactly matches the correct answer, as well as partial credit for getting some characters right. If an evaluation run gets interrupted partway through, the scripts can resume from where they left off using saved checkpoint files. This project is aimed at researchers working in computer vision and robotics who want to reproduce the paper's results or evaluate their own models against the same benchmarks. Running it requires access to powerful graphics hardware and a machine learning model already prepared to be tested, since the scripts are built around evaluating an existing model rather than teaching how to build one from scratch.