robot-i-o/rio

RIO is a Python tool that lets researchers control many different kinds of robot arms, robot grippers, cameras, and remote control devices through one shared interface, instead of writing separate code for each brand of hardware. The project supports robot arms from several manufacturers, including Franka, Kinova, Universal Robots, UFACTORY, and SO100, so a single codebase can work across different robots without being rewritten. It is built for the field of robot learning, where researchers collect data from real robots, sometimes by having a human operator control the robot remotely, called teleoperation, and use that data to train AI systems. RIO includes built-in support for collecting this kind of data, for teleoperation itself, and for running Vision-Language-Action policies, a type of AI model that takes in what a robot sees and a written instruction, then decides how the robot should move. Grouping all of this behind one interface means a lab that switches from one robot arm brand to another does not need to rewrite its data collection or control code from scratch. The project depends on a companion library called openpi from Physical Intelligence, which must be downloaded separately into the project folder. Setup has been tested on Ubuntu 22.04, optionally with a real time kernel patch installed for more precise timing, and uses the uv tool to manage the Python environment and install dependencies. Documentation can be built and viewed locally using mkdocs. The README itself is quite short and mainly covers installation steps rather than usage details, so most information about how to actually run data collection, teleoperation, or policy deployment with RIO lives in the project's separate documentation site rather than the main page. Since the project is aimed at robot learning researchers rather than general developers, using it in practice assumes access to compatible physical robot hardware and familiarity with that research area.