1. What can we learn from touch? : towards end-to-end learning from robot touch
- Author
-
Church, Alex D. J. R., Lloyd, John, Lepora, Nathan, and Santos-Rodriguez, Raul
- Abstract
Physical interaction is fundamental to the human experience. Our embodiment has been optimised through aeons of evolution for the ability of general purpose physical interaction. Currently, there is a distinct gap between the physical capabilities of humans and those of robots. In robotics, rich contact-based interaction is often minimised due to the complexity that arises from these interactions. The information carried in these interactions is difficult to perceive and to model, leading to a lack of physical skill in robotics. In humans, touch is a fundamental sense for enabling our physical capabilities. This work examines the use of touch in robotics for perception and manipulation, aiming to improve the physical skills of robots. The goal of this work is to combine recent advances in tactile sensing technology with powerful and scalable robot-learning methods. To this end we successfully apply several learning methods to data obtained with a TacTip tactile sensor. We examine the use of deep learning for tactile perception, with both a single sensor and multi-fingered robotic hand. We apply end-to-end deep reinforcement learning, directly from sensory inputs to control outputs, for a complex tactile task requiring sensitive perception and precise action. We provide a scalable and efficient method of simulating tactile data, demonstrating zero-shot sim-to-real transfer of multiple learned tactile polices. Finally, we make progress towards solving a grand challenge of in-hand dexterous manipulation, using a combination of the above approaches. The successful application of these techniques provides evidence that physically interactive tasks can be solved by focussing on them through a tactile lens. The application of general and scalable techniques such as deep reinforcement learning shows potential, as these methods progress many new tasks will become achievable goals. Overall, we make progress towards our overarching aim of endowing general purpose robots with the ability to physically interact in our world.
- Published
- 2023