Emlyn’s research interests include, computer vision, machine learning, and autonomous vehicles.
- AgriFoRwArdS CDT Annual Conference (2022): Inverse Reinforcement Learning Applied to the Correspondence Problem.
- The Towards Autonomous Robots and Systems (TAROS) Conference 2023 / CDT Annual Conference / Joint Robotics CDT Conference (September 2023): Simulation of Collective Bernoulli-Ball System for Characterizing Dynamic Self-stability.
- Represented the CDT at the Douglas Bomford Trust bi-annual meeting (March 2022).
- British Science Week (School Outreach) demonstrations and stall (March 2022).
My name is Emlyn and I’m from Anglesey in North Wales, before joining the CDT I was studying mechanical engineering at the University of Leeds. In my spare time I enjoy sailing, pub quizzes, playing guitar and watching films.
I was looking forward to joining the CDT and am excited to learn more about robotics and artificial intelligence before choosing my project. I chose the CDT because I think that increasing autonomy in agriculture is an important step in increasing food security. Coming from an area where most of the land is used for agriculture, I am interested in learning how that land could be used more efficiently.
Inverse Reinforcement Learning Applied to the Correspondence Problem
In Learning from Demonstration (LfD), the correspondence (or cross embodiment) problem is the problem of mapping between a demonstrator and a dissimilar learner. This project aims to explore how Inverse Reinforcement Learning (IRL) can be used to learn a reward function that is independent of the embodiment performing the demonstration, and to use this reward function to teach a real robot to perform a task.
Robot Learning of Harvesting Tasks from Humans’ Demonstrations
Agricultural tasks, such as selective harvesting and crop husbandry, are usually physically demanding, challenging, tedious and repetitive which are performed by humans in harsh conditions . The project aims to increase the automation level of such tasks by reducing its cycle time and rise the harvesting success rate focusing on complex and challenging task variants. E.g. When a specific fruit must be collected from a cluster. Therefore, during the project, an intelligent robotic system will be developed which will be able to learn the harvesting task directly from human experts and autonomously adapt the acquired knowledge to novel scenarios.
The project will build upon and expand current state of the art methods both in the field of robot learning and manipulation. The robotic system is expected to learn how to perform agricultural tasks from demonstrations which will be provided by observing human while they perform the task. The collected data will include arms’ trajectory, RGBD sensory input and tactile information. The harvesting policies will be formulated as nonlinear time-invariant dynamical systems whose parameters will be optimized through machine learning methods and will be fine-tuned from the interaction of the robotic system with its environment using reinforcement learning methods. Furthermore, the use of dynamical systems will allow the coordination of the individual systems of the robot such as both arms and the grippers through coupling methods. Such approach allows the execution of the task in reduced time and the development of compliant controllers which will be able to compensate external perturbations
The student will acquire skills on the fields of machine learning, computer vision and control theory specially focused on supervised learning and dynamical systems. Furthermore, will be able to identify challenges rising from applying machine learning methods for the control of embodied systems such as robots and identify suitable machine learning methods which can deal with those. During the project, the student will acquire hands on experience on developing machine learning and control methods and applying them on a robotic system. Additionally, the student will develop critical evaluation skills required for designing, performing, and assessing scientific experiments. The outcomes of the project will be published to highly influential international journals and conferences in the fileds of robotics and machine learning. Additional training will be available through participation in workshops, Ph.D. level courses and summer schools.