The Robotics Interview Is Moving Back On Site
By Chris Escobar • 5 minute read
Something is shifting in how robotics companies interview, and it has less to do with robotics than with a verification problem the whole industry is having.
The take-home assignment is on its way out. Not gone — Karat’s 2026 survey still has 45% of employers using them and 63% running automated code tests — but the direction is set, and the reason is boring. Nobody can tell who wrote the submission. That same survey found 22% of candidates using AI to generate answers during a live interview, which should tell you what’s happening to the asynchronous version. When you can’t verify authorship, the artifact stops carrying information.
So the evaluation is moving back into the room. Live coding is up. Meta added an AI-enabled round in late 2025 and is expanding it this year, where the thing being scored isn’t whether you can write the function but how you prompt for it, how you review what comes back, whether you catch the model’s mistakes, and whether you own the result. Other employers have gone the surveillance route, adding eye tracking and screen-share requirements, which seems like an expensive fix for a problem better solved by asking a decent follow-up question.
Robotics watches this from a slightly odd angle, because robotics never had a clean proxy to begin with.
A perception stack that works in sim and falls apart on a wet warehouse floor is a familiar disappointment to anyone who has shipped a robot. This field has a ground truth that doesn’t negotiate. You cannot prompt your way past a motor that stalls under load, or a timing bug that only appears when the arm is carrying eleven kilos instead of five. The move everyone else is scrambling to make, from artifact to observed behavior, is one robotics teams were already halfway through.
The lab day is doing the work now
The pattern happening at small companies: a short, real bench task. An eval board — an STM32 Discovery, an RP2040 — and an ask like write a driver for this sensor, handle the interrupt, now find the bug on the scope while I watch. It isn’t a puzzle. It’s twenty minutes of the actual job, and it’s close to unfakeable.
There’s a second-order effect worth noticing. Reports from hardware hiring suggest companies that include an onsite day see higher offer-acceptance rates. Read that carefully. The day that exists so the company can evaluate the candidate is also the day the candidate decides they want in. You can’t fall in love with a robot over Zoom. If you’re a founder still running an all-remote loop to save two thousand dollars in flights, you may be paying for it at the offer stage without knowing.
What founders are actually screening for
Here is where candidates most often misread the room, especially candidates coming from big companies into a 25-person robotics startup.
The technical bar is table stakes, and the founder usually establishes it in the first fifteen minutes. Everything after that is a different test. Can you make a decision with half the information you’d like? Will you tell the founder they’re wrong, in the room, in front of people? Have you ever had a robot in a customer’s building, at 2 a.m., failing?
That last one carries disproportionate weight right now, and it’s the one people underrate on their résumés. A team of strong software engineers with no deployment experience will spend six months learning what one operations-scarred person knows on day one. Founders have been burned by this and are screening hard for it. If you’ve done field deployment, real hours and real sites and real failure modes, that belongs near the top of your story instead of buried under the framework list.
The other signal founders name repeatedly is curiosity that runs past the assigned task. The candidate who asks why the requirement exists, or what happens downstream of their piece, reads as someone who will still be useful when the roadmap changes in November. Which it will.
The backdrop
Demand is genuinely strong. Market trackers put growth in physical-AI roles somewhere around 35%, and there aren’t enough senior robotics engineers in the U.S. to fully staff the programs that are already funded. Take the market-size projections with the usual grain of salt, since the people publishing them are rarely disinterested, but the shortage at the senior end matches what I see week to week.
That shortage gives candidates room to negotiate. It does not soften the loop. A tight market makes companies move faster and pay more; it does not make them lower the bar on judgment, because at 25 people a bad senior hire is a visible, expensive, morale-denting event that everyone watches happen.
What I actually want to know
All of the above is the view from outside the room. I hear about these processes secondhand, from founders explaining why someone didn’t work out and from candidates explaining what they think went wrong. Both accounts are edited, and both are edited in the teller’s favor.
What I don’t have enough of is the version from people who just went through it and are now three months into the job. The moment in the loop that surprised them. The question that was harder than expected. The thing they’d tell themselves at the start.
If you’ve been hired into a robotics or physical AI company in the last year or so, I’d like to hear it. Send me a DM on LinkedIn. I’ll write up what I learn, and no names get attached to anything unless you want them there.
Sources: Karat, Engineering Interview Trends 2026; IEEE-USA, Three Ways AI is Reshaping Traditional Technical Interviews; Recruiting From Scratch, Hiring Embedded Systems Engineers at a Hardware Startup; a16z, The 7 Hires a Hardware Startup Needs to Get Right.