The Year I Taught a Robot to Hesitate

I handed in my master's dissertation this year. The title is a mouthful , Safe Tool Handover in Human-Robot Collaboration: A Dual-Stream Framework Combining VLA Failure Detection and Human Intent Recognition , but the question underneath it is simple enough to explain to my daughter.

When a robot hands you a tool, how does it know you're ready to take it?

This is the story of how I spent a year answering that, while running a startup and trying to be present at home. If you want the science, I wrote a technical deep-dive and you can read the whole thing. This post is the other half , the part that doesn't make it into an abstract.

Where it started

I kept coming back to one uncomfortable fact. We have robots now that can listen to a spoken instruction, look at a cluttered table, and pick up a tool they've never seen before. Vision-Language-Action models made that real. And yet every single commercial deployment I could find still keeps those robots behind a fence, or working a shift when the humans have gone home.

The machines are capable. We just don't trust them next to us.

The more I read, the more I understood why. When one of these models meets something it wasn't trained on , a shadow falling the wrong way, a tool that's a little too long , it doesn't stop and ask for help. It keeps going, confidently, toward the wrong action. In a warehouse with no people around, that's a dropped box. In a handover, with a hand reaching in, that's an injury.

The thing I wanted to build wasn't a smarter robot. It was a robot that could hesitate. One that watches itself closely enough to notice when it's about to get something wrong, and watches you closely enough to know whether you're actually ready. That instinct to hesitate at the right moment , that was the whole project.

The constraint I kept forgetting

Here is the part I underestimated: I had one year.

The MSc at BSBI is built to be done in a single year, and I said yes to that with the confidence of someone who hadn't done it yet. A year sounds like a lot until you subtract a full-time job. In October I became CTO and co-founder of Alma, building AI systems for a problem that genuinely matters and does not wait for anyone's coursework. The dissertation had to fit into the edges of that.

So the first real decision wasn't technical. It was honest. I could not run this study on a physical robot. Hardware, ethics approval, human participants reaching toward a machine that might fail , none of that fits in a year, and none of it is responsible to rush. I built the whole thing in simulation instead, and I spent a real amount of the dissertation defending that choice rather than apologising for it. Simulation has the lowest external validity and the highest internal control, and for a question about which safety architecture is better, control is exactly what you need. It's also the only ethical way to deliberately make a robot fail near a human-shaped object a thousand times over.

Naming your limits clearly turns out to be a research skill, not a weakness. My supervisor taught me that.

On having a good supervisor

Professor Vincent English read this dissertation chapter by chapter, and every round came back sharper than I'd handed it in. He had three questions he asked relentlessly. Compare things directly instead of describing them one after another. Tie every number back to the question it was supposed to answer. Make each chapter lead into the next instead of just stopping.

It sounds basic written down. In practice it reorganised how I think. I came in wanting to report , here is what I did, here is what happened. He kept pushing me toward argue , here is the claim, here is the evidence that could have killed it, here is why it survived. The pre-registered criterion in my methodology, the thing I'm proudest of structurally, came directly from that pressure. I committed in writing to what would count as success before I collected any data, so the result could fail in plain sight. That's his fingerprint on the work.

The struggle nobody writes in the abstract

There were weeks the simulation simply wouldn't behave. The Franka Panda arm couldn't reliably pinch the thinnest tool at the floor of its reach, and I lost days to it before making a deliberate call to model the grasp as a fixed attachment and study the monitoring layer instead of the mechanics of gripping. Writing that decision down honestly , as a caveat that shapes how every later number should be read , was harder than making it.

And there was the quieter struggle. A single-year master's on top of a startup means the dissertation lives in early mornings and late nights, in the hours that would otherwise belong to the people you love.

My wife, Nicklah, carried more of those hours than I can repay. She gave up a great deal of our time together and never once made me feel guilty for it. She also proofread these pages with an eye I'd lost , she caught things I had read so many times I could no longer see them. My daughter was patient with a father who was often at his desk, and honestly, the sight of her was what pulled me back up on the days the work felt pointless. My parents taught me, a long time ago, how to stay steady when things get difficult and how to finish what you start. That advice held me together more than once this year.

I dedicated the dissertation to the three of them, and to my parents. That dedication is the truest sentence in the whole document.

What I found

The short version: it worked, with the honesty the work demanded.

The full system , watching the robot and the human at the same time , caught 81.5% of the failure trials, where a robot with no monitoring caught none and finished every dangerous handover anyway. It never once let a failing grip reach the hand. And it did this without slowing down the handovers that were going to succeed fine. Safety and smoothness, it turns out, didn't have to be traded against each other.

But the finding I respect most is the one that didn't flatter me. The two halves of the system catch different kinds of failure. Watching the robot is nearly perfect at catching the robot's own mistakes and almost useless at catching a human reaching in at the wrong moment. Watching the human is the only thing that helps there , and even then it only lifted that number from 8% to 28%. The human-side stream is the weak part of my own design. Saying so plainly, instead of burying it, is part of the contribution. It tells whoever picks this up next exactly where to dig.

That's the lesson the whole experiment was built to teach, and it's bigger than one handover task: a robot isn't safe or unsafe in the abstract. It's safe or unsafe doing a particular thing next to a particular person , and the only way to know which is to watch both.

Coming up for air

Finishing this didn't feel like a fireworks moment. It felt like quietly closing a laptop at the end of a very long year and going to find my family.

I started because I wanted to understand why capable robots are still kept away from us. I finished understanding it better, with a small, careful piece of evidence about how to close that gap, and an even clearer sense of what I still don't know. That's about the most a year of research can honestly give you, and I'll take it.

If you want the architecture and the numbers, the technical write-up is here. If you want the whole thing, read or download the dissertation. And if you're in the middle of your own one-year-and-a-job version of this: name your limits early, find a supervisor who makes you argue instead of report, and thank the people holding the rest of your life together. Out loud. In writing. While it still counts.