Researchers build missing infrastructure to move AI between robots

Carnegie Mellon University researchers developed Robot I/O, or RIO, an open-source software framework intended to reduce the extensive setup work needed before robot-learning experiments can begin. The framework supplies a unified interface for robot control, data collection, teleoperation and AI deployment, aiming to let users transfer software components, data and AI systems among robotic arms, humanoids and other platforms.

The project addresses a common problem: laboratories often create custom software for each machine, so code must be rewritten when hardware changes. That limits sharing across groups and can hinder reproducibility. RIO uses modular, reusable components that researchers can combine and adapt rather than reconstructing basic infrastructure for every robot or project. The team says this approach can support more consistent experiments and make robotics development easier for newcomers.

In a setup test, undergraduate intern Reya Shukla, who had machine-learning experience but no robotics background, unpacked a robotic arm and configured it for teleoperation in roughly two hours by following RIO documentation. Researchers contrasted this with the weeks or months, and sometimes much longer, commonly spent getting new robots ready for research.

Team members also used RIO while developing policies for humanoid and bimanual robots. Its swappable building blocks allow the same pipeline to be used across platforms without special code for data collection or policy training, including systems with extra cameras or arms. The researchers argue that shared infrastructure is especially important for gathering robot data and supporting newer general-purpose AI control methods. RIO remains an active research project. Future work includes broader hardware support and reducing the effort needed to bring robots online. Some contributors are extending related technology through Lavoro AI, a startup co-founded by Jean Oh, with a longer-term goal of robotics foundation models that adapt across tasks, environments and platforms.