General Profile:

I first worked with AI in my summer project for MSc Computer Science, where I was able to implement modern computer vision as well as the more traditional MiniMax algorithm. Through this project and subsequent study, I nurtured an interest in AI that would drive me towards the Interactive AI CDT here at Bristol.

I believe in the power of AI to create positive outcomes, but that this technology must be managed appropriately if we are to realise the benefits. The emphasis on ethical behaviour and human-in-the-loop systems make this CDT stand out as a place to develop AI research and implementation skills alongside ethical processes.

As a novice to AI, I consider a training year essential. The CDT stands out for it's variety of academic content, and while I'm not settled on a direction for my future research, I look forward to the perspective this variety will afford.

Research Project Summary:

My research as a PhD student in the field of interactive AI and robotics revolves around improving the capability of robots to learn from human demonstrations. Presently, robots necessitate an extensive number of precisely orchestrated professional demonstrations to execute a specific task, and their ability to generalise from such demonstrations is limited. My focus is on empowering non-professionals to teach robots how to perform tasks with only a minimal number of demonstrations.

To achieve this goal, I am delving into imitation learning through behavioural cloning and inverse reinforcement learning. Within these areas, I will investigate what methods are being used to record and learn from human demonstrations with an aim to improve upon them throughout my PhD. Recent innovations in human body and hand pose estimation enable more accurate representations of human demonstration, where the human's actions from visual observations can be directly translated into a 3D simulated environment. This improves significantly on previous methods where accurate recording of task demonstration required teleoperation or physical manipulation of a given robot - both of which require expert intervention.

Building on this, I am investigating how these demonstrations can be encoded into goals and sub-goals for learning effective robot policies. By combining existing techniques in observation embeddings through unsupervised learning, evolutionary updating of design-specific control paradigms, Imitation/reinforcement learning and Sim2Real, I endeavour to reduce the amount of human effort required to train an effective, generalisable robot policy.

Currently, my work primarily involves an extensive review of the existing literature and gaining a comprehensive understanding of the robotics domain. In the near future, I aim to identify specific areas where I can make valuable contributions. This includes conducting experiments, working with robotics hardware and software, and publishing research papers that not only contribute to the field but also enhance my practical expertise.

Looking ahead, my overarching goal is to synthesise existing methodologies in novel ways to reduce the number of demonstrations required for training a robot in a new task. This research direction aligns closely with the EPSRC's AI research area.