Nobody's Checking the Numbers (Or the Brakes)

On April 22, on Tesla's Q1 2026 earnings call, Elon Musk told investors that Optimus production at Fremont would begin "somewhere around the late July, August timeframe." By the time Q2 rolled around, their shareholder update read production is "anticipated... later this year."

The strangest thing about humanoid robot production is that almost nobody has to prove the numbers. But to be fair to Elon, he did caution that the Optimus production rate itself was "literally impossible to predict." And sure enough, the manufacturing table from Tesla's own Q2 2026 shareholder update, published July 22, listed the Fremont and Texas Optimus lines as still under "Construction." The same table already lists Cybercab and Megapack as "Production." Tesla didn’t just miss the original target; it went on to replace it with a vaguer one. That's as close to a self-administered test as this industry gets, and Tesla failed it in public.

AgiBot and Figure create a different problem. On June 28, AgiBot announced that its 15,000th robot had rolled off the line. On April 29, Figure announced that BotQ had delivered more than 350 Figure 03s and had demonstrated a production cycle of one robot per hour. Multiple outlets like Interesting Engineering reported the Figure numbers, but the underlying figures came from Figure itself. In both cases, the public is being asked to accept a company-reported count. There is no audited production tally, serial-number ledger, or third-party inspection evident in the sources covering either claim.

That does not make either number false. It makes them difficult to verify. The same would be true of any startup reporting a factory total: without an outside check, the reader has little way to distinguish a production record from a production claim. The number may be right; but the evidence is still thin. A claim can be detailed without being independently verified. Figure gives us a rate and a cumulative count; AgiBot gives us a milestone number. More detail is not the same thing as more evidence. Neither company made a promise precise enough to create the kind of before-and-after check Tesla accidentally created for itself.

This leaves humanoid robotics in an odd evidentiary position. Production numbers are being used to tell us who is scaling, but there is no shared public standard for proving what “scaling” actually means. Tesla got caught because it supplied a deadline. AgiBot and Figure supplied numbers. One claim failed a test; two claims have not really faced one.

So when the next 10,000-robot milestone arrives, I will be double checking the source.

The safest way to stop a walking humanoid may not be to stop it too quickly.

That sounds backwards because, for most industrial machinery, faster shutdown is the point. Cut the power, the machine stops doing things, and the hazard goes away. The Siemens researchers behind a new study show why that logic breaks when the machine is balancing on two legs: removing power from a walking humanoid causes it to fall. The fall is now part of the safety problem.

The researchers found that a humanoid's safe stopping behavior depends on its posture and where it is in a step. A stop commanded while both feet are down is a different problem from one commanded while one foot is swinging forward (single-support phase). During that single-support phase, the robot may need to finish the step before it can regain a stable stance.

I checked the paper's own timing table, and the measured mechanical stopping time ranged from 0.3 to 1.0 seconds on the Unitree G1 they tested. That’s because a bipedal robot stops faster when both its feet are planted, slower when caught mid-stride. But that range is not the story. The key finding here is that “faster” is not automatically safer. The researchers describe an overly aggressive halt as capable of pushing the robot outside the region from which it can recover its balance, turning the protective action into the thing that causes the fall.

The paper says existing machinery standards can govern the detection and signaling parts of a stop, but not the robot's own reaction once the stop command reaches its balancing controller. Agility's Digit has already cleared a safety inspection at one site, for detection and signaling, not the reaction chain this paper describes. There is no humanoid-specific standard governing that reaction chain, and the researchers explicitly do not claim end-to-end safety certification for their system. Even the usual calculation for how far away a safety barrier should sit depends on a stopping time that changes with gait phase and remains uncertified.

So the industry has inherited a safety instinct built for machines that become safer when they become inert. Walking machines cannot become inert without potentially becoming dangerous. Until there is a defensible lower bound for how fast a humanoid can safely stop, what exactly does “emergency stop” mean when the emergency is happening halfway through a step?

Both stories this week come down to the same gap: an industry running on trust nobody's actually earned. Tesla is the only company that made a claim precise enough to fail a real test; AgiBot and Figure's numbers were never built to be tested in the first place. The humanoid emergency stop is the same problem from another angle — an instinct borrowed from machines that don't walk, never checked against the ones that do.