Imitation learning for contact-rich manipulation

Currently, our goal is to extend the CTF IL framework [6] to handle contact-rich tasks by
improving the way object interaction occurs. The objective is to create a method that accounts
for force feedback as well as being closed-loop in order to adapt to errors arising during contact
or from the visual feedback that can occur when approaching the object. Initially, we are focusing
on scenarios where a new demonstration is provided for each new object, where providing a
representative demonstration to solve the task is possible. Then, in the future we aim to explore
methods that build control policies based on 3D object models constructed directly from vision and
tasks that require tactile feedback to be solved. Additionally, we aim to extend this to be able to
generalise demonstrations across similar objects and utilise prior knowledge obtained either in the
real-world or simulation to provide solutions for tasks were providing a clear enough demonstration
is not possible. Looking forward, the ambition is that by the end we can have a framework where
humans can teach robots to solve complex, contact-rich manipulation tasks in the real-world, with
only a handful of demonstrations and improve from interactive human feedback.

General research area: robotics