My research lives at the boundary of learning-based robotics, sim-to-real transfer, and scalable ML systems. I care most about work that eventually runs on real hardware in unstructured environments.
Six interconnected themes that have guided my research agenda since 2019. Every project touches at least two of them.
Bridging the physics gap between simulation and the real world is arguably the central challenge in robotic learning today. My work investigates domain randomisation schedules, residual physics learning, and adaptive system identification to make policies robust to unmodelled dynamics.
Deep RL has transformed what robots can do in simulation; translating those results to physical systems requires rethinking reward design, curriculum scheduling, and constraint satisfaction. I focus on making RL algorithms sample-efficient enough for real hardware loops.
Grasping and in-hand manipulation in unstructured environments remain unsolved at the precision required for industrial deployment. My work combines imitation learning, contact-rich simulation, and tactile sensing to push success rates on novel objects above the 90% threshold.
Quadruped robots operating in the real world must traverse terrain that is never fully characterised in advance. My work on terrain-aware policy conditioning and periodic gait controllers has shown that learned locomotion can match model-based approaches on rough outdoor terrain.
Vision is the primary sense for most robotic manipulation pipelines, yet deploying vision models in industrial environments with poor lighting, occlusion, and domain shift remains a persistent challenge. My focus is on efficient architectures for edge deployment and self-supervised adaptation.
Learning-based methods complement, but rarely replace, classical planning and control when safety guarantees are required. I investigate how learned heuristics can accelerate sampling-based planners and how MPC frameworks can incorporate learned uncertainty models for robust trajectory optimisation.
Research projects in various stages of progress — from data collection through to paper submission. Each is a collaboration with researchers across Europe and North America.
We are equipping a UR5e arm with a custom tactile sensor array and investigating how force-torque signals can close the loop on sub-millimetre peg-in-hole tasks that are currently beyond the precision of vision-only policies. Data collection is ongoing at the ITU Robotics Lab.
A systematic empirical study investigating what fraction of sim-to-real transfer failures can be attributed to (a) inaccurate physics parameters, (b) sensor simulation errors, and (c) distribution shift in visual inputs. We are running paired experiments across Isaac Gym and MuJoCo against physical UR5e trials to produce a variance decomposition across failure modes.
Following our ICRA 2023 work on efficient EfficientDet variants, this project investigates whether self-supervised contrastive pre-training on unlabelled factory floor imagery can reduce the labelled data requirement for defect detection by an order of magnitude. Paper currently in writing; targeting IEEE RA-L submission by Q3 2026.
We are investigating whether a quadruped locomotion policy conditioned on a learned latent terrain prior — estimated online from proprioceptive history alone — can generalise zero-shot to terrain categories unseen during training. Building on prior work with ANYmal-C; current experiments running in Isaac Gym with ETH ASL.
Research is a team sport. These are the researchers I collaborate with most actively across Europe and beyond.
My work has appeared at the leading robotics and machine learning conferences and journals.
"The goal of research is not to produce papers — it is to understand something well enough that you could build a system that works. Papers are the record of that understanding. If the system can't be built, the understanding isn't complete."
If you're working on problems that overlap with my research areas, I'm always interested in hearing about potential collaborations — especially ones that involve real hardware.