Embodied robot learning

Most robots have rigid bodies and are made of mostly hard components. Often they are limited to specific tasks, rather than
solving general problems. They are not autonomous, very energy inefficient and dangerous to work with. In order to
overcome this limitation the concept of soft robotics gained tractions. It opens up a whole new dimension in robotics
research: it allows robots to be flexible, to be safe and to interact actively with their environment [1]. Since then, research
field has been rapidly growing, providing new technologies for sensing, actuation and adaptation. At the same time,
machine learning algorithms have developed due to the cheaper and more powerful computational processing, however,
they are still not efficient enough for robotic applications due to the large gap between hardware architectures and software
components. The aim of this PhD is to bring these technologies closer to each other and to develop new methods for robot
learning.
The goals of this project are 1. to investigate and to develop new soft robotic sensing and actuation techniques, 2. to develop
new machine learning algorithms and 3. to bring together the hardware components and software algorithms into one
embodied design.
Firstly, the significant engineering challenges in this project will be to characterize, model, fabricate and test a range of soft
sensors and actuators, such as soft fluidic devices, smart material structures and microfluidic devices. They will need to be
characterized in order to understand their behavior and optimized with respect to their sensing and actuating performance.
This characterization will be carried out using electrical and optical measurement devices at Bristol Robotics Laboratory.
Secondly, machine learning algorithms will be investigated and will be chosen according to their potential for performing
high level robotic tasks, such as grasping, locomotion or adaptation to external environment.
Thirdly, we will investigate two possible ways of the integration of the soft robotic components with machine learning
methods. In the first case, learning algorithms will be combined with in-body and in-silico features. The aim of this part of
the project is to find new ways to couple existing and new machine learning methods with soft robotic devices. In the other
case, the learning algorithms will be fully implemented into the soft robotic body and the computation will not require any
additional computing unit. This method will have the advantage of being compact: for example, the deformation of the body
will have an impact on the learning process, therefore, our robots will be able to act according to their environment. This
idea is closely related to the concept of morphological computation, which understands morphological features as a way to
embody computation. As a result complex computational tasks (including learning to adapt) can be outsourced to the
morphology of, e.g., a robot.
In summary, this project is aimed to investigate the implementation of the learning processes directly in robotic bodies, and
to have a significant impact on the fields of robot learning, bioinspired robotics and soft robotics.
[1] - Kim, S., Laschi, C. & Trimmer, B. Soft robotics: a bioinspired evolution in robotics. Trends in biotechnology 31,
287-294 (2013)
[2] - Hauser, H., Ijspeert, A., Fuchslin, R., Pfeifer, R. & Maass, W. Towards a theoretical foundation for morphological
computation with compliant bodies, Biological Cybernetics, Springer Berlin / Heidelberg, 105, 355-37 (2011)