Research
Robot learning and embodied AI: methods that let robots acquire, adapt and generalise complex behaviour, and that can run on real robotic systems.
Current research
World models for generalisable robotic manipulation
Robot learning can now produce manipulation skills that generalise across tasks, but transferring them across embodiments, to robots with different physical and dynamic constraints, remains hard. My thesis studies whether embodiment-agnostic world models can be used to plan for embodiment-specific robots, taking an aerial manipulator performing contact-rich manipulation as the case study.
Aerial manipulation is a demanding test: contact and arm motion affect the flight dynamics, the platforms are typically underactuated, and demonstrations are expensive to collect. That makes it a good setting to ask how much manipulation knowledge can be reused across very different robots.
Intermediate results and blog posts will be linked here as they are published.
Broader interests
- Deep reinforcement learning: model-free and model-based RL, policy optimisation, exploration, robust skill acquisition.
- World models and representation learning: predictive dynamics, latent representations, generalisation across environments.
- Generalisable robot learning: cross-task and cross-embodiment transfer, imitation learning, VLAs, robotic foundation models.
- Learning-based planning and control: combining learned policies and predictive models with optimisation and model-based controllers.