Deployment is running ahead of autonomy, and the gap between the two is being filled by people. That is my reading rather than a reported fact, so what follows is the evidence behind it - and the clearest of it isn’t in any demo video. It’s on the hiring pages of the companies building the robots.

The disclosure nobody reads

Figure advertised a role called Humanoid Robot Operator in San Jose at $25–35 an hour. The duties are worth reading as a technical document: “Be responsible for the humanoid robot operating successfully for the Customer use case,” “Run the robot constantly throughout the day, identifying bugs and problems,” and the line that matters - “May wear teleoperation equipment and guide the robot through designated behaviors.”

RiVR advertised a Remote Robot Supervisor in Austin to “operate robots in remote locations, including customer sites, test environments, or field deployments.”

Both listings have since closed. The pages are still up, which is the point - these were public the whole time.

Treat them as disclosures rather than ads. Read plainly, they say a robot in front of a customer had a person assigned to it, running it all day, ready to put on teleoperation equipment when the policy ran out of road. None of that is scandalous, and my read is that it is how first deployments of anything have always worked. Two listings are also not a survey, and I can’t tell you what the other companies do. But these are specific, checkable claims about the current state of autonomy, published by the companies themselves, and that is more than any highlight reel offers.

Autonomy is real, in a narrow band

The opposite evidence deserves equal weight, because it exists.

Epoch AI’s February 2026 survey of deployed robot capabilities reports a DYNA-1 run that “completed a 24-hour continuous run folding 850+ napkins at roughly 60% of human speed with a 99.4% success rate and zero human interventions,” and a commercial laundry deployment that “folded over 200,000 towels across 10 commercial clients in three months, with a 99% quality acceptance rate.” Those are vendor-reported figures rather than independent audits, but they are specific enough to be wrong, which is more than most of this field offers.

The same report is blunt about the other end. Amazon has “deployed over a million robotic units across its facilities, though most have limited autonomy,” and one vendor “claims to have run eight-hour shifts in campuses and corporate settings for over six months,” though “many edge cases required teleoperation.”

So the honest summary is not “robots can’t do it.” It’s that autonomy holds inside a narrow band - one task, one environment, tight tolerances - and the width of that band, not the peak capability, is what a buyer is purchasing.

The ratio is a restatement of one number

The operator-to-robot ratio gets discussed as a staffing choice. It isn’t. It’s arithmetic on how often the robot needs help.

Take an operator on an eight-hour shift: 480 minutes of attention to spend, none of it lost to breaks. Assume an intervention costs ten minutes of that attention - notice, take control, recover, hand back.

That ten minutes is an assumption. I could find no public figure for it, and every number below rests on it, so it is worth saying plainly where it came from: it is a short remote intervention, not a walk to the machine and a bug report afterwards. Along with the no-breaks assumption above, it errs in the robot’s favour. If the real figures are worse, every conclusion below gets stronger rather than weaker.

• One operator to ten robots leaves 48 interventions per shift, or 4.8 per robot. That is one intervention every hundred minutes of robot runtime.
• One to fifty leaves about one intervention per robot per shift - roughly one every eight hours.

That second number is the ask. A humanoid doing varied physical work needs to run an entire shift with a single human touch before one supervisor can watch fifty of them. Set against the compounding arithmetic in my last post - where a twenty-step task at 95% per-step reliability fails about two attempts in three - the distance is not incremental. Halve the assumption to five minutes and the requirement halves with it, to a little over four hours, which is the bottom line on the chart. That is a friendlier number and still most of a shift.

The research I can find treats this as an open problem rather than a solved one. The authors of ARMADA describe existing human-in-the-loop deployments as systems that “usually require full-time human surveillance during policy rollout,” and their contribution is a “greater than 2× reduction in human intervention rate” - real progress, measured against a baseline of someone watching constantly.

What it costs

Take the bottom of Figure’s posted range, $25 an hour, on the same principle.

At one operator per robot, that is $25 per robot-hour in posted wages. At 1:10 it falls to $2.50. At 1:50, $0.50.

Read those as floors rather than costs. A posted wage is not what an employer pays: benefits and payroll overhead sit on top, and I have no public figure for either at these companies. Nor does any of it count capital, maintenance, or the onsite staff who still handle what a remote operator can’t.

Now scale it. A hundred-robot fleet at 1:10 needs ten operator seats per shift. If the fleet only runs weekday days, which is how the Figure role is actually posted at 40 hours a week, that is ten people and nothing more. Run it around the clock, which is the whole point of buying one, and each seat takes 168 hours a week against that same 40: 4.2 people per seat, forty-two in total. I have not added anything for leave or attrition, so read forty-two as a floor too.

Either way it is a department, growing linearly with the fleet, inside a company whose entire pitch is that labour scales sublinearly.

That is the inflection to watch, and it lands well before the fleet looks impressive. As I argued in the first post, one supervisor to ten machines is a labour-arbitrage business and one-to-three merely relocates the labour. The job ads are the earliest public evidence I know of for where a given company actually sits, and they appear well before any of it reaches a financial statement.

Where I could be wrong

My read is that the intervention rate is falling, and the papers above are the reason to expect it to keep falling: both are about extracting more from each human touch. If post-training drives it down fast enough, the operator pool never becomes a department - it stays a team, and the linear-scaling problem never arrives.

My ten-minute assumption may be the wrong shape rather than merely the wrong size. If interventions can be queued and handled asynchronously - the robot pauses safely, waits, and an operator clears a backlog - then attention stops being exclusive and the ratios improve sharply without autonomy improving at all. Failure detection is what gates this, and ARMADA’s detector reports “nearly 95% accuracy on average” at spotting failures autonomously.

And human intervention isn’t free quality. ROVE notes that intervention trajectories are “often suboptimal,” and that methods treating them as expert supervision “can absorb hesitant, inefficient, or even erroneous behaviors.” The operator in the loop is not a clean fallback; they are another source of variance.

All of it would show up in one disclosure: interventions per robot-hour, over a stated window. Until someone publishes it, the hiring page is the better document.