Research Programme 2026–2027

What this page is

This is a research programme, not a research result. It exists so that the direction of my work is legible before there is a paper to point at, and so that the projects, notes, and experiments published on this site can be read against a stated plan rather than assembled after the fact.

Nothing on this page reports a finding. When findings exist, they will appear as their own research entries with methods, metrics, failure analysis, and code.

Starting position

I come to this from software engineering rather than from a research group. That background is genuinely useful — building and operating non-trivial systems is most of what an experimental pipeline is — but it does not substitute for the mathematical and methodological preparation that research requires. The first phase of this programme is therefore explicitly about closing that gap in public, with artefacts rather than assertions.

Concretely, that means implementing things from first principles rather than importing them: estimators before filters libraries, optimisation before training loops, geometry before pose libraries.

Direction

I am interested in agents that act in the physical world, and specifically in the join between perception, semantics, and control:

• Embodied AI — how an agent's representation of a scene should be shaped by the fact that it has to move through it.
• Robot learning — learning policies and representations for navigation, manipulation, and control, and understanding when learning actually beats a well-specified classical method.
• Computer vision — visual perception and representation as the input side of that loop.
• Autonomous systems — sensing, estimation, planning, and decision-making as a system rather than as separate components.
• Multimodal intelligence — combining vision, language, sensor observations, and action, and being honest about the computational cost of doing so.

These are interests, not claims of expertise, and they are expected to narrow. The purpose of the programme is to make them narrow for a defensible reason.

Method of working

The working method is deliberately conservative, because the failure mode I want to avoid is producing demonstrations that look like results.

1. Foundations before questions. Linear algebra, probability, optimisation, estimation, and rigid-body geometry, each with implementations I wrote and tests I can point to.
2. Reproduce before extending. Re-run an established baseline and match its reported numbers before proposing a change to it. A reproduction that fails is itself worth publishing.
3. Controlled experiments. Fixed seeds, recorded configurations, a stated hypothesis per run, and metrics defined before the run rather than after it.
4. Negative results stay. Runs that disprove the hypothesis are documented, not deleted. That is most of the reason to keep a research log at all.
5. Everything traceable. Each claim on this site should lead to a repository, a configuration, and a command.

What will appear here

As the programme progresses, this site should accumulate:

• foundational projects with real implementations and tests;
• research notes and derivations under Writing;
• reproduction studies of published baselines;
• a research log documenting the progression, including the parts that did not work;
• and eventually, one or two focused research projects with their own methods, experiments, and limitations.

Current state

Early. The site is being published at the start of the programme rather than at the end of it, which means it is sparse by design. I would rather it be verifiably thin now than impressively vague.

If you are a researcher working on embodied or multimodal autonomy and any of this overlaps with your own questions, I would genuinely like to hear from you — contact details are here.