What I work on

On-road autonomy leans hard on structure: lane geometry, sign conventions, prior maps, a distribution of scenes that repeats. Off-road removes almost all of it. Terrain is unlabelled and self-similar, traversability is a property of the surface rather than the paint on it, and appearance shifts with season, weather and time of day faster than any fixed training set can cover.

My thesis investigates diffusion models as a way into that problem. Their appeal is that they learn a distribution over plausible scenes rather than a single deterministic mapping — a natural fit for a setting where the honest answer is often a set of possibilities with different likelihoods rather than one confident prediction.

Research on explainable AI. The part of the problem I find most interesting is the gap between the two audiences an explanation has to serve: the evaluation metric, which wants faithfulness to the model, and the human being, who wants a reason they can act on. Those are not the same target, and a method can score well on one while being useless for the other.

Where someone looks is a remarkably strong signal about what they are doing, and it is available before they act. This work classified human activity from gaze, which is a good problem precisely because the input is so sparse: a scanpath carries far less information than a video frame, but it sits much closer to intent.

A HoloLens 2 supplied the eye tracking, through a capture application I built after the obvious off-the-shelf route turned out to sample too slowly to be useful. Volunteers performed four ordinary industrial activities in the DFKI smart factory — walking, waiting, using a tool, assembling. Walking and waiting are the awkward pair, and they were chosen for exactly that: nobody is looking at anything in particular, so the gaze is at its least structured.

Features drawn from the blinks and the fixations fed a set of classifiers. The more instructive result is the one that failed: the recurrent model I added to the inherited methods lost, and kept losing. The dataset is small, which is not what such a model wants, and manufacturing more of it synthetically made things worse rather than better.

A four-person project lab, and a full perception-to-manipulation loop on real hardware: find a cuboid placed anywhere on the bench, recover its pose, and have a seven-axis arm pick it up and set it down elsewhere without touching the obstacles it already knows about.

Pose is the interesting half. A custom-trained YOLO detector finds the cuboid in the colour frame; corners taken from inside that box are solved against a CAD model of the object with PnP, and the estimate is steadied over several frames before the planner is allowed to act on it. Motion planning runs through MoveIt, with fallbacks for the poses a single planner cannot reach — which, for this task, is often.

The instructive part of a project like this is never the detector — it is everything between a bounding box and a gripper that closes in the right place.

Motion control for FUME, a reconfigurable lower-limb exoskeleton for paraplegic users built in the lab. Alongside it, technical documentation for a 6R industrial manipulator.

The inertia of a machine like that you can write down. The Coriolis and gravity terms, with a person strapped into it, you cannot — so the controller learns that part online instead: adaptive tracking built on a generalised fuzzy hyperbolic model, chosen because it is cheap enough to run in real time on a plant that is nonlinear, multi-input and permanently disturbed. It ran on the robot itself rather than in simulation, and outperformed a tuned PID.

This is the work that made me a researcher: a control problem where the plant is a person, the failure modes are not abstract, and the specification is written in terms of what a human body can tolerate. It produced my first publication, below.

The classical hard case in control: more degrees of freedom than actuators, an unstable equilibrium, and no way to cheat. I built and compared three controllers on it — linear-quadratic, classical PID, and fuzzy — which is the most efficient way I know to learn what each family of controller actually buys you.