Abstract
In tactile sensing for robotics the main goal has been to create robots able to complete tasks quickly and with accuracy – both very important metrics. But little has been done to address more practical factors such as the training time, amount of data needed for training and time collecting it, transferability to other hardware or robustness to changes in tasks or environment. These may be just as, if not more, important factors that need to be considered for the wide spread adoption of tactile sensing in robots.In this thesis we seek to address these less considered metrics and develop online learning methods with the aim to overcome the challenges presented by data-intensive, offline, deep learning methods. We implement contour following of planar objects with a single tactile sensor on a robot arm and compare performance under different conditions with different stimuli. It is shown that data-efficient methods can be just as accurate as deep learning methods but without the hours of data collection and pre-training that deep learning methods require, and can be much more adaptable to different tactile stimuli.
We then show these online learning methods can be transferred to new robotic platforms with ease. This is done by developing the first reported instance of a high resolution tactile foot for a quadrupedal walking robot, and demonstrating it following the edges of raised paths using only tactile feedback in the feet as guidance. The robot is able to learn where the edge is in order to avoid falling off the path, guiding it safely across this challenging terrain.
Finally, we look towards being able to follow contours in 3D, increasing the applicability to real world objects (that are often not planar) but also increasing the dimensionality of the problem. It is implemented on a robot arm with a single sensor. We show that the methods still retain their data-efficiency over deep learning methods and their robustness to novel stimuli, demonstrating that online learning may be key to enabling robots to cope with the complex and ever changing real world.
Date of Award | 23 Jan 2024 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | David A W Barton (Supervisor) & Nathan F Lepora (Supervisor) |