UKRI AI CDT in Dependable and Deployable AI for Robotics (CDT-D2AIR)

Calum Arnott

Project
Learning and Control for Contact-Rich Robotic Manipulation

Summary
Robotics has the potential to assist humans with complex physical tasks, particularly in areas such as assistive living and by taking on dull, dirty, and dangerous work. However, reliably interacting with the real world remains a significant challenge in robotics, in part because of uncertainty in system dynamics, noisy sensing, and the complexity of contact interactions. My research explores how machine learning and model-based control can be combined to enable safe, adaptable, and reliable robotic behaviour. Machine learning enables robots to adapt to complex environments through interaction data, while model-based approaches provide structure, interpretability, and the ability to enforce safety and constraints. What interests me is how the complementary strengths of these two sets of tools can be integrated to allow robots to learn from experience while maintaining dependable performance in tasks involving physical interaction. My goal is to enable robots to carry out tasks with greater autonomy by improving how they make decisions under uncertainty, ultimately contributing to robotic systems that can operate safely and effectively in the real world.