The Soft Hand Dexterous Learning project will develop data efficient learning methods optimized for soft systems. We will iteratively co-develop a series of soft grippers which are well suited for learning along with a library of learning methods which are suitable for self-supervised learning with soft grippers. This work will build on our prior success rapidly learning mobility strategies for soft systems to enable near real-time learning of dexterous skills without simulation or demonstration.

A robot hand with tactile sensors will be used to test active-perception methods for differentiating surface features.