What is imac?

iMaC is a research project from a 2026 academic paper that teaches a computer how to predict what will happen when a robot arm tries to pick up or move an object. The core idea is to translate a robot action command into two kinds of images: one showing how things should move, and one showing where contact with surfaces will occur. Those images then guide a video-generation model that produces a short clip of the predicted outcome. Think of it as a simulator that does not need to run physics calculations in real time. Instead of computing every force and collision, the system generates a video of what the scene should look like after the robot acts. That video can then be used to judge whether a policy (the set of rules telling the robot what to do) is working correctly, without having to run the real robot for every test. The code covers the full research pipeline: preparing training data from a dataset released for the CVPR 2026 World Model competition, training the model in two stages, and running evaluations either offline against pre-recorded episodes or online against a live simulator. Three separate workflows are included: a first training stage using replay images only, a second stage that adds 3D scene information, and a WorldArena track that targets a different benchmark. Setting up the project requires a machine with a compatible GPU, Python 3.11, and a collection of supporting libraries including PyTorch, a diffusion-model toolkit, depth-estimation tools, and robotics utilities. The dataset and pretrained model weights are downloaded separately via provided scripts. Environment variables point the code to your local copies of those files, so no paths are hardcoded. This repository is aimed at robotics researchers working on world models and embodied AI. There is no graphical interface and no built-in visualization beyond the video outputs the inference scripts produce. If you are not already working in this research area, the setup and concepts will require substantial background reading.