flâneur

Training an SO-101 RL Agent to Grasp and Lift in MuJoCo

ggando.com · 1,591 words · saved by 1 readers

When I attended a robotics meetup in SF this summer, I realized that the era of robotics is on the rise, and now is a good time to start building robots. I've been working on imitation learning for my SO-101 robot arm setup, successfully training two ACT policies via LeRobot: green cube → white paper, and red cube → green bowl. Both worked… sometimes. Maybe 50% success rate on a good day. This is the same problem as supervised deep learning, where every improvement requires more data in the form of manual demonstrations. I obviously need more demos to go from 50% to 80% success; it could improve with 50+ more episodes. But I'd need to keep collecting data every time I change the situation or object. Big corporations like Google and OpenAI have teams collecting millions of demonstrations, and I'm seeing the same situation happening in Tokyo. As a solo developer, I can't really compete with IL-based methods. So I decided to go the RL route. Ideally, I train a good model in simulation fir

Published on 2026-01-02 Context # When I attended a robotics meetup in SF this summer, I realized that the era of robotics is on the rise, and now is a good time to start building robots. I've been working on imitation learning for my SO-101 robot arm setup, successfully training two ACT policies via LeRobot: green cube → white paper, and red cube → green bowl. Both worked… sometimes. Maybe 50% success rate on a good day. This is the same problem as supervised deep learning, where every improvement requires more data in the form of manual demonstrations. I obviously need more demos to go…

saved by

related reading