What is foca?

FOCA is the code release for a robotics research paper presented at ICML 2026, focused on helping robots learn new physical skills from only a small number of example demonstrations. Today's vision-language-action models, which combine seeing, understanding language, and producing robot actions, tend to perform poorly when there is not much training data available for a new task. FOCA addresses this by training the robot to reason about likely future outcomes of its actions rather than just copying the exact motions in the demonstrations it was shown. The method works by combining a prediction of future interactions with the environment and an alignment step that checks whether the robot's plan matches those anticipated future goals. This future-oriented approach lets the model make better decisions over longer sequences of actions, and it also allows the system to train partly on synthetic video generated by separate world models, without needing real robot action labels for that extra video data. The paper reports state-of-the-art results on both simulated benchmarks and real-world robot manipulation tasks. The repository provides code to reproduce experiments on two simulation benchmarks, LIBERO and RoboCasa, along with pretrained model checkpoints trained on different amounts of data, from 10 percent up to the full dataset, all hosted on Hugging Face. It documents the installation steps, how to run evaluation on the simulators, how to fine-tune the model on a new dataset, and how to co-train using a separate video world model called DreamGen. It also includes guidance for adapting the method to a user's own robot data, along with configuration flags that turn on or off specific parts of the FOCA method such as the future-object tracking loss. This project is aimed at robotics and machine learning researchers working on vision-language-action models who want to reproduce the FOCA paper's results or build on its approach to learning from limited demonstration data.