About Solarian
Solarian is an open-source Python framework for building AI-powered robots, and a home for the open hardware and models that go with it. It was started in May 2025 by Raul Garreta, after selling his second company, MonkeyLearn, to work on something closer to a life's work.
Why another robotics framework
Before building Solarian, Raul spent a long time with the existing tools. ROS 2 felt low-level, hard to set up (especially on a Mac), and designed for an earlier era of explicitly programmed control. LeRobot did a great job making modern robot learning and affordable open-source arms accessible, but its architecture wasn't the foundation he wanted to build on.
The biggest frustration was reuse. Each project started over: collect demonstrations for one robot, train a policy for that embodiment, and start from scratch when the arm, the gripper, or the camera setup changed. Fixed third-person cameras also tied every setup to one workspace, which doesn't work for robots that move around.
Built around UMI
The answer came from UMI, the Universal Manipulation Interface from Stanford, Columbia, and Toyota Research Institute. UMI policies describe the gripper's motion rather than one arm's joints, so the same policy can run on different compatible arms without retraining. A camera travels on the gripper, so the view stays consistent across robots and places. And you collect demonstrations by hand, with a portable handheld gripper, instead of teleoperating a robot through a leader arm.
Getting UMI to work reliably took months. Solarian adopts it as its main approach to learning manipulation, and does the engineering work to make it practical for other developers. Work on visual prompting, pointing at the exact object the policy should act on, grew out of the same effort.
Why it still makes sense in 2026
- AI agents need building blocks too. Reimplementing a robotics stack from scratch wastes time and tokens, and Solarian has been tested on real robots, where things break and hardware is never fully reliable.
- A framework is a set of opinions. Its architecture, conventions, and trade-offs come from experience and taste, which are hard for an agent to invent on its own.
- Developer experience still matters. Clear conventions, good documentation, and a coherent architecture help humans and AI agents alike.
- It's more than software. Solarian also hosts open hardware: a handheld data collector built on a Raspberry Pi with a camera and IMU, an actuated gripper that costs under $100 to build, and a six-DOF modification of the SO-100 arm for UMI-style manipulation.