A key reason for a lack of widespread adoption of soft, bio-inspired and otherwise unconventional robots is the difficulty of controlling them. While we can rely on expert-crafted dynamics and kinematics models for conventional robots, which are rigid and precision manufactured, such models are difficult to obtain and calibrate for unconventional robots. This project investigates approaches to learning models of any robot self-supervised.

A 3D differential kinematics model learned of a robot hand.