My PhD work asks how robots can learn precise physical skills, use limited demonstrations effectively, and generalise beyond a rehearsed setup. The emphasis is on questions that connect learning methods with the realities of contact, perception, control, and physical systems.
Dexterous, contact-rich manipulation
How robots can learn precise manipulation skills when success depends on force, contact, timing, and object geometry.
dexterous manipulation · contact-rich manipulation · physical intelligence
This is the central application area for my current PhD research: manipulation problems where sensing, action, and physical interaction cannot be separated cleanly.
Questions
- How can robots learn contact-rich skills from a small number of demonstrations?
- How can those skills generalise reliably to new objects and real-world situations?
Imitation learning and generalisation
How robots can acquire new behaviours from demonstrations without requiring extensive task-specific retraining.
robot learning · imitation learning · in-context learning
I am interested in methods that make demonstrations useful beyond a single rehearsed setup, especially when the robot must adapt to variation in objects or scenes.
Questions
- What should a robot infer from a small set of demonstrations?
- How can in-context imitation learning adapt behaviour while retaining precision?
Representations and perception for manipulation
How representations of objects, scenes, and physical interaction can support robust manipulation decisions.
perception for manipulation · tactile sensing · generative modelling
This direction connects computer vision and tactile sensing with the learned representations that manipulation policies use to choose and refine actions.
Questions
- Which visual or tactile information matters for a manipulation skill?
- How can generative models represent variation in physical interactions?
Reinforcement learning and control
How learning-based methods and explicit control structure can work together in reliable embodied systems.
reinforcement learning · control systems · embodied AI
My project work in value learning, estimation, and embedded control provides practical testbeds for examining assumptions, failure modes, and measurable behaviour.
Questions
- Where does reinforcement learning add value alongside modelling and feedback control?
- How should learned behaviours be evaluated when they run on physical systems?
These are active directions rather than claims of completed results. Related engineering work and experiments are documented under Projects and Writing.