🙏 This was a super exciting collaboration! Huge thanks to Uksang Yoo (@uksang.bsky.social), Jean Oh, Christopher Atkeson and Jeffrey Ichnowski (@jeff-ichnowski.bsky.social) for their invaluable guidance and support. soft-spin.github.io.
5/5🧵.

🦾 Using CMA-ES (Covariance Matrix Adaptation Evolution Strategy), the robot could quickly identify effective movements using only the real robot (no simulation). With as few as 10 generations, the soft robot learns how to spin various objects.🤲💡

🔄 The robot practice consistently using a soft hand attached to a manipulator arm. To make the initial state repeatable, we place dropped pens into a fixture, and the robot picks it up and tries again. ⏪🔧

🎯 Surprisingly, we do not need the robot to plan motion every step. By representing the entire pen spinning motion with just 8 variables, we reduced the action space, making it easier for the robot to optimize spin actions in the real world. ⚡️🖊️
2/5🧵.

Can soft robots rapidly spin pens like humans?🤔 We’ve shown that soft robot hands can master the dynamic tasks of pen spinning—no hours of GPU training or complex sim-to-real needed! Check out soft-spin.github.io. 🤖✍️ @cmurobotics.bsky.social .
1/5🧵.