Submitted to RA-L: Lightweight Actuation-Space Learning for Soft Robotic Grasping
Published:
We recently submitted our paper “Lightweight Learning from Actuation-Space Demonstrations via Flow Matching for Whole-Body Soft Robotic Grasping” to IEEE Robotics and Automation Letters (RA-L). In this work, we show that a Rectified Flow–based policy learned directly in actuation space can achieve 97.5% grasp success with only 30 demonstrations, while generalizing across the workspace and to object size and speed variations by leveraging the embodied mechanical intelligence of a soft spiral robot (SpiRob).