Every humanoid launch arrives with a video, and the video is an advertisement. That is not an accusation. The job of a launch clip is to show the best thing the machine has ever done, in the best conditions the team could arrange, cut down to the best thirty seconds. The trouble is that most of us then reason about the robot as if we had watched it work, when what we watched was a claim. My beat at Robot Brief is evidence, so here is how I turn a claim back into a question, and the four things I check before I let a demo move me.
A demo is a hypothesis, not a result
Start with the right posture. A demo is the seller's best case, which makes it a hypothesis about what the machine can do, not a record of what it does. The burden of proof sits with the company, not with you, and an unproven hypothesis is worth close to nothing until someone tests it. That sounds harsh. It is the only stance that has kept me from being wrong in public, because the failures in this field almost never happen in the clip. They happen in the ninety-nine takes you did not see, and in the long months between the launch and the ship date.
So I watch a demo the way an auditor reads a number: assume nothing, and ask what would have to be true for the impressive part to hold up in a normal room, on a normal day, with no team standing off camera. Four questions do most of that work.
Question one: is it driving itself?
The single most important thing a demo usually hides is whether the robot was autonomous or teleoperated. A person in a motion-capture suit, or a skilled operator with a game controller, can make a machine look graceful, deliberate, and smart, and none of that grace belongs to the robot. It belongs to the human hidden behind it.
Teleoperation is not cheating. It is how a lot of these systems collect training data, and a teleoperated robot doing real work in a warehouse can be useful today. The dishonesty is in the edit, when a company shows a teleoperated clip and lets you assume autonomy. So I look for the tell. Does the company say, in plain words, that the behavior was autonomous, and for how long? Does the motion carry the small hesitations and corrections of a machine deciding for itself, or the confident flow of a human hand? If nobody states it, I assume there was a hand on the strings and discount the clip. A company proud of true autonomy tends to say so loudly, because it is the hardest thing to fake and the most valuable thing to own.
Question two: how many takes did that cost?
The second question is about reliability, and reliability is a number, not a highlight. One clean run proves the machine can do the task once, under the right conditions, with the crew ready. It says almost nothing about whether it can do the task the next hundred times.
The figure that matters is the success rate across many attempts in unscripted conditions, and that figure almost never survives the edit. A robot that folds a shirt beautifully in the launch video might fold it correctly six times out of ten on your actual laundry, which is a party trick, not an appliance. When a company gives you a real number, a stated success rate over a stated number of trials, I take them seriously, because they have shown me the thing that is easy to hide and painful to publish. When they give me one flawless take and a swelling soundtrack, I treat the capability as unproven and move on.
It is also worth asking about the room. Demos happen on flat floors, under even light, with known objects placed at known spots. Your home has none of those courtesies. A machine tuned to a staged set can fall apart the moment the lighting changes or a chair moves six inches, and the clip will never tell you how brittle that tuning is.
Question three: can you buy it, and for how much?
The third question is the most boring and the most clarifying. Can you order the machine today, at a stated price, with a delivery date? A launch with no price and no ship date is a concept. A product has both.
Price is where a lot of impressive robots quietly die. A machine can work and still be a failure if it costs more than the labor it offsets, or more than a household would ever pay for a helper that sometimes drops a plate. Watch what a company does with its number. A confident team names a price and holds it. A nervous team says "contact us," then raises the figure after launch as the true cost of hands and actuators comes into view. The price tells you whether the engineering problem and the manufacturing problem have both been solved, and the second one is harder than the first.
Question four: who checked, besides the company?
The fourth question is about independence. Every claim in a launch comes from the party with the most to gain from you believing it, so I ask who else has touched the machine. Has an outside reporter run it, unscripted, in a room the company did not prepare? Has a paying customer put it to work and said, on the record, that it did the job? Has anyone with no stake in the outcome verified the number?
This is where I watch for a specific kind of laundering, where a company cites forecasts and press mentions as if they were evidence about the machine. A market projection is not a product review. A wall of logos from publications that "covered" the launch tells you the public relations worked, not that the robot does. Independent, hands-on testing by someone who wants to find the flaw is the only thing that turns a demo from a hypothesis into a fact, and it is exactly the thing a weak product avoids.
The forecast trap
One trap deserves its own section, because it fools careful people. It is the slide from a real event to a projected market.
Take two claims that circulated the same week. Virtuix sold its first enterprise motion system to Tesla's Optimus group. Separately, a forecast pegged the companion-robot market at $1.1 billion by 2030. These feel like the same kind of statement, and they are not alike at all. The first is a transaction you can check: a supplier, an order, a customer, a thing that already happened. The second is a spreadsheet about a year that has not arrived, built on assumptions you cannot see. One is evidence. The other is a guess in the costume of evidence, and the decimal point is part of the costume. I keep those in separate columns, and I never let a 2030 projection do the work a shipped order should have to do.
Why consumers should care most
If you follow robotics as a spectator, sloppy reading costs you nothing. If you might actually buy one of these machines, it costs you real money and possibly your safety. The distance between the demo and the product is the distance between the robot you were sold and the robot that shows up at your door, and consumers stand at the far end of it, after the enterprise pilots and the press cycle have moved on.
So the method is not academic. Before you preorder a home robot on the strength of a beautiful video, ask the four questions. Was it driving itself. How often does it succeed. What does it cost and when does it ship. Who checked besides the seller. If the answers are not on the table, you are not buying a product. You are funding a hypothesis, and you deserve to know that is what you are doing.
The words that should make you suspicious
Marketing language leaks. Once you know the phrases companies reach for when the truth is thinner than the video, you can read a press release like an x-ray. "Up to" is the loudest one. "Up to forty tasks" means one task forty times in a good week, or forty tasks once each on a lucky day, and never all of them at once, reliably. "Designed to" and "capable of" describe intent, not performance, and a machine capable of climbing stairs may manage it once in ten tries. "In ideal conditions" is a confession in a lab coat. "Representative footage" in small type at the bottom of a video means what you are watching did not happen in a single take. None of these phrases are lies. They are careful, and careful language around a product usually protects a number the company would rather you not work out.
Reading one demo, start to finish
Put the method together on a single clip. A company posts a humanoid unloading a dishwasher, set to music, one continuous shot. Question one: does the caption say autonomous? If not, assume an operator, and notice the motion is a touch too smooth, a touch too certain, for a machine solving the problem live. Question two: is there any number? One shot means one success, and a dishwasher full of your grandmother's glasses is a place where a six-in-ten success rate is a catastrophe. Question three: is there a price and a date, or only a link to join a waitlist? Question four: has anyone outside the company loaded that dishwasher, in a kitchen the company did not build? If the answers are no, silence, waitlist, and no, then you have watched a hypothesis with a soundtrack, and the right amount to update your beliefs is close to zero. That is not pessimism. It is arithmetic.
What an honest demo shows
For contrast, here is what a demo looks like when a company is not hiding. It shows more than one take, including a failure or two, because a team sure of its success rate is not scared of the misses. It states, in text on the screen, whether the run was autonomous and how long it lasted. It puts a price and a ship date somewhere in the announcement. And it points to someone outside the company, a reporter or an early customer, who ran the machine with no team standing by. Demos like that are rare, and when I see one I move a long way toward belief, because the company has volunteered the exact evidence a weak product spends all its energy avoiding. The willingness to show the misses is the signal. The rest is production values.
What lands
Here is the rule I keep coming back to. A demo shifts the burden of proof onto the seller, and it stays there until someone independent, holding no stock, watches the machine work in a room nobody prepared. Until that happens, the right response to even the most stunning clip is a flat and friendly "show me the success rate, the price, and the ship date."
None of this makes me a cynic. I want these machines to be real, and some of them will be. It makes me an auditor, and an auditor's favorite sentence is a number you let a stranger check. The next time a robot goes viral, do not ask whether it looked amazing. It did. Ask what it would take to prove it, and then wait for that, not the video.