Robotics, perception & autonomy

Infrastructure fails slowly and in public, and we still mostly find out by sending someone to look. I want to work on autonomous inspection: systems that reconstruct a structure's geometry, locate damage inside that model, and decide where to look next.

I have built pieces of this pipeline already. Glance & Glamour turns a photograph into a 3D body mesh: the generative model is TripoSG, and my part is the platform around it — the inference service and CUDA container that serve it, and the browser preview that lets you orbit the result. Copernican integrates physical motion against a live external feed and renders it in the browser. WatchWay closes the far end, taking hazard reports as GIS telemetry and ranking repairs by severity through a priority queue.

What I want from a research group is the join: learned damage detection sitting between the reconstruction and the dispatch, so that a digital twin is an assessment of a structure rather than a picture of one, and the inspection decides its own next move.

The physical training is not decorative here. A crack in concrete and a failing hydraulic line are materials problems before they are computer-vision problems, and my degree is in the mechanics of physical systems rather than in images of them.

Planning for machines in the real world is rarely one solver problem. It is several competences that have to agree, under constraints that are physical rather than logical. I want to formalise how specialised agents divide a plan and stay consistent when the world pushes back.

Locus is the working version: seven specialised agents coordinating pathfinding, travel cost, safety, weather and air quality into a single plan in real time. EagleSight is the rigorous version: vehicle routing with time windows as a mixed-integer program, minimising fuel against hard scheduling constraints.

The question I actually want to answer sits between them. When the optimiser's constraints come from a physical model that is itself uncertain — a fluid-mechanics estimate of whether an engine will survive the shift — how much should the planner trust it, and what does it do when it should not?

The binding constraint on food security is not how much is grown but how much survives storage. Detecting grain quality changing in real time, without destroying the sample, is the same perception problem wearing different clothes.

NIR and gas-sensor telemetry is the input; the interpretation layer is the same deep-learning and autonomous-dispatch work as the tracks above. The optimisation target is energy — drying and storage that hold quality without wasting heat.

I keep this track because it is where my degree is strongest and because biosystems and bioresource departments run some of the best applied robotics work there is. If a group's autonomy problem happens to live in a field or a grain store, I am already fluent in the domain.