Three robot jobs that exist in 2026, and three that don't

Humanoid robot videos are everywhere, and they make it look like the robots are ready to clock in. They mostly aren't, and the reason is simple: the video is not the market. Most of the robots doing human jobs today are dull, narrow and measurable, and the dull ones are the ones that ship. I read the trade press so you don't have to. This issue is about which robot tasks are actually earning their keep, and which ones are still a demo with a press release attached.

What happened

Two recent stories tell the real version. The Robot Report covered Groceryshop 2026, where retail robots looked ready to scale. A month earlier, IEEE Spectrum took a hard look at humanoid safety and economics. Read together, they say the same thing: the deployments that work are narrow.

The gap is generalization. An impressive demo does not prove a robot's behaviour generalizes, as Appen's Jeanine Sinanan-Singh explained on the Robot Report podcast. A robot that does one tidy run in a lab has told you almost nothing about a Tuesday afternoon in a real building.

I think that is the story of 2026. The physical AI that ships looks like a shelf scanner, not a butler.

Three robot jobs that exist

Shelf scanning in stores

Simbe has rolled out 3,000 Tally robots globally across retailers and countries, and an ongoing rollout at a Tier 1 retailer will grow that number. Tally drives the aisles and scans shelves. Badger Technologies, with Jabil manufacturing scale behind it, now calls itself a retail intelligence company rather than a robotics company, and is exploring estimates of shelf quantity, replenishment timing and freshness. ShelfOptix goes after smaller shops: a foldable robot escorted between drug, convenience, dollar and regional grocery stores by a human, instead of one robot per store.

The honest limit: this is a data job, not a handling job. The robot looks and reports. It does not fix the shelf.

Moving totes and boxes in warehouses and factories

Agility Robotics has more than US$300 million in multi-year customer orders for Digit 5 across manufacturing, warehousing and logistics. The robot lifts 23 kg up to 2.1 m, works near people without physical barriers, and runs more than 20 hours a day. That is tote and box work, on a known route, all day.

The honest limit: IEEE estimates those orders amount to under 1,000 robots, one tenth of Agility's RoboFab capacity in Oregon. An order book is not an installed fleet.

Long straight welds in shipyards

Persona AI builds a humanoid for shipyard welding, starting with long linear welds. A ship needs hundreds of kilometers of them, and welding speed is capped at about a centimeter per second. That slow, millimeter-scale, repeated motion is exactly what a machine does well. It has public partnerships with HD Hyundai and POSCO, both in Korea, and says it wants customers who can support hundreds of robots per site.

The honest limit: the pitch is more ships built, not lower costs, because its partners are labor-constrained with a backlog. The company still needs tens of thousands of hours of expert demonstration data and real-world testing.

Three that don't (yet)

Picking products off store shelves for customers

Blue Collar Robotics, a startup pitch finalist, is developing a mobile robot to physically pick products in the store, moving items between shelves and carts. Initial customer trials are planned. The reporter covering it calls the problem exponentially more difficult than shelf scanning. I think that is generous to everyone's timelines.

The drop-in humanoid that replaces a worker one-for-one

Agility's illustrative estimate puts Digit 5 offered as a service at about US$8,500 a month, and suggests that at 20 hours a day against a human cost of US$30.50 an hour it could save an employer about US$100,000 a year per robot. IEEE points out the assumption underneath: that the robot is a drop-in replacement for human labor, which the writer doubts has ever happened with any robot. I agree. Treat that figure as a ceiling, not a forecast.

The factory line that retools itself in minutes

Industrial Next, founded in 2021 by ex-Tesla engineers Lukas Pankau and Allen Pan, builds AI robotic stations it says can be retooled in minutes instead of weeks without traditional programming. It announced a partnership with BMW Group Technology Office to evaluate its dexterous manipulation. Hold onto the word evaluate. Today, changing a line usually means reprogramming and retooling over weeks or months, and robotics still lacks the large datasets that precise physical work needs.

How mature is this

My read: the robots doing human jobs that last are boring on purpose. Greenfield sites designed from scratch favour specialised non-humanoid robots. Brownfield sites were built for humans, with stairs, narrow aisles and doorways, and that is where a human-like form starts to pay off.

The real test is the second purchase order, not the demo. Fleet deployment needs at least 10,000 cycles per shift with near-zero drift, and deployments fail on drift, sensor fatigue and recalibration rather than on reasoning. Nobody posts a video about recalibration.

What it means for a business

Start from the task, not the robot. Pick a job with economic value and work backwards. If a vendor opens with the robot, ask which task it was designed for and whether that task exists in your building this quarter.

Then ask for four metrics, because these are the ones finance teams actually care about:

• Mean time between interventions, which is the least often published.
• Cost per unit of work: per pick, per meter, per inspection.
• Redeployment time to a new task.
• Safety incidents per operating hour.

On renting versus buying, the one public number is Agility's illustrative estimate of about US$8,500 a month for Digit 5 as a service. That is the company's own rounded estimate, not a market price, and it assumes drop-in labor replacement, which is the part I do not believe yet. Renting does move the risk to the vendor, which is where it belongs while the metrics above are still unpublished.

What to watch

• Agility's SPAC merger, expected to close within months with more than US$620 million raised, plus EU and UK availability in 2027. That tells you whether production scaling is real.
• The BMW Group Technology Office evaluation with Industrial Next. An evaluation is not a rollout, so watch for it turning into a deployment.
• Blue Collar Robotics trials and the Simbe rollout at a Tier 1 retailer. Both test whether the harder store work and the bigger fleet actually land.

Get the next issue every Tuesday: subscribe to the Briefing.