Over the past few weeks, I built the SO-ARM100, an open-source robotic arm developed by The Robot Studio in collaboration with Hugging Face. The SO-ARM100 (and its successor, the SO-101) is a low-cost, 3D-printable leader-follower robotic arm system designed to work with the open-source LeRobot library, making it an accessible platform for hands-on experimentation with robotic manipulation and, eventually, machine learning.
I wanted a project that would let me get my hands dirty with the full stack of a robotics build, mechanical assembly, actuation, and eventually software and AI, rather than just reading about it. The SO-ARM100 turned out to be the right fit: well-documented, actively maintained, and designed from the ground up to be paired with modern robot learning tools.
Responsibilities:
- Sourced all mechanical and electrical components for the build, including STS3215 servos, motor control boards, power supplies, and cabling
- 3D printed the arm’s structural parts, calibrating printer settings (layer height, infill, supports) to meet the tolerances required for properly fitting servo housings and joints
- Verified print accuracy using the project’s gauge system before committing to full part production, to avoid wasted prints and assembly issues down the line
- Assembled the follower arm from printed parts and off-the-shelf components, including wiring and mounting the servos and motor control board
- Set up and debugged the servo motors, using the LeRobot library (and manufacturer debugging tools where needed) to configure and test each joint
- Researched the broader LeRobot ecosystem to understand how the arm integrates with teleoperation and future machine learning workflows
Outcomes:
- Successfully built a fully assembled, functional SO-ARM100 robotic arm from raw parts and printed components
- Gained practical, end-to-end experience spanning mechanical design tolerances, 3D printing calibration, electromechanical assembly, and low-level motor control, skills that connect directly to my interest in robotics
- Developed a working understanding of the LeRobot framework, positioning the arm for future software integration rather than treating it as a standalone mechanical exercise
- Identified my next phase of work: integrating AI-based training models so the arm can learn to pick up objects using computer vision, moving the project from teleoperated hardware toward an autonomous, vision-guided manipulation system
What’s Next:
The immediate goal is to pair the arm with a vision system and train a model to perform object pick-and-place autonomously. This means moving beyond manual teleoperation and into the realm of learned policies, likely starting with imitation learning approaches supported by LeRobot, then layering in a camera-based perception pipeline so the arm can identify, localize, and grasp objects without direct human control. I’ll be documenting that process in future posts as the project evolves from a mechanical build into a working AI-driven robotic system.