A humanoid doing backflips "looks incredible and tells you almost nothing about whether it can do real work in a real facility," Agility Robotics posted last month. That takes an entire ecosystem, they said.
Days later, the same account posted a job: a Staff Robotics Software Engineer to build "the teleoperation platform that teaches Digit how to interact with the physical world." Low-latency human demonstration data collection, "at production scale."
Read those two posts together and the ecosystem behind the anti-backflip robot comes into focus. It is a human wearing a control rig.
Last week the humanoid debate was about whether viral clips prove anything. This week Agility itself supplied the sharper frame — "there's a difference between a pilot and a partnership" — and in doing so handed us the exact yardstick the industry keeps refusing to pick up.
Because here is the trick the "anti-demo" turn is playing. It denounces one unfalsifiable spectacle, the acrobatic clip, and quietly installs another: teleoperation-collected "data" and self-reported success rates. Both are demos. One just looks like humility.
It denounces one unfalsifiable spectacle, the acrobatic clip, and quietly installs another: teleoperation-collected 'data' and self-reported success rates. Both are demos. One just looks like humility.
The backflip at least tells you the truth about what it is — a stunt. The teleoperation platform tells you a robot is "learning to interact with the physical world" while a person is doing the interacting.
Watch how the numbers get built. AGIBOT's G2 robots reportedly completed 64,828 tasks at a 99.99% success rate during a six-day livestream at Longcheer Technology's factory. That figure sounds like an autonomy benchmark.
It is a livestream metric — curated conditions, a fixed window, a chosen line, tasks like tablet inspection and material transport. What it does not report is how many of those 64,828 tasks touched a human hand, remote or otherwise, and how often someone intervened.
Avatar Robotics gives the same shape of claim without the ambiguity. Its mobile manipulators have "processed and assisted in shipping more than 900,000 products" in live customer facilities, including a global beauty retailer.
Assisted is the load-bearing word. Avatar's own model is explicit: remote human operators control the robots to pick, pack, kit, sort, and count today, and the resulting "control and task data" feeds foundation models "over time."
Its investors are refreshingly blunt about it. AlleyCorp's Brannon Jones praised the human-in-the-loop system as "performing at a level competitive with skilled human operators." Of course it is. A skilled human is operating it.
So "900,000 products" is not a measure of what the robot can do. It measures how much work humans did through the robot — reframed as machine progress and a training dataset.
This is the maneuver the MIT Technology Review documented in detail: the human labor gets hidden, and the public then "overestimates the machines' actual capabilities."
The scenes are specific and strange. A worker in Shanghai spent a week in a VR headset and an exoskeleton, opening and closing a microwave door hundreds of times a day, to train the robot standing next to him.
1X's Neo, a $20,000 home humanoid shipping this year, is not committed to any level of autonomy. When it gets stuck or faces a tricky task, a teleoperator in Palo Alto pilots it — looking through the robot's cameras to iron your clothes or unload your dishwasher.
MIT Tech Review names the honest description of that arrangement: not autonomy, but "wage arbitrage that re-creates the dynamics of gig work," now for physical tasks, performed wherever labor is cheapest.
None of this makes the robots vaporware. The human workforces are real and doing real work. The problem is what happens when they stay invisible — you mistake concealed labor for intelligence, and you see far more autonomy than exists.
Agility's "pilot versus partnership" line is exactly the right instinct. It's just pointed at the wrong target. The distinction that matters isn't between a demo and a deployment; it's between a robot a human is driving and a robot that increasingly drives itself.
And there is one number that would settle it. Not products shipped, not tasks logged, not a success rate. The intervention rate — how often a human has to take over — and what this article treats as its inverse, the fleet-per-operator ratio.
If teleoperation is genuinely teaching autonomy, that curve bends. One operator supervises two robots, then five, then twenty, and the takeovers per hour fall. That is what "collecting data to improve autonomy over time" is supposed to produce.
No humanoid maker has published that curve. Not Agility, not Avatar, not AGIBOT, not 1X. They publish the top-line spectacle — 900,000 products, 99.99%, 64,828 tasks — and keep the denominator that would expose the human labor firmly offstage.
Avatar's pitch literally promises "single operators to manage larger robot fleets." Good. Show the ratio, month over month. Until then, the promise is indistinguishable from a business that scales by hiring more remote pilots.
The uncomfortable comparison is a company most humanoid press never mentions, because it isn't humanoid at all.
Brain Corp reportedly surpassed 50,000 BrainOS-powered robots deployed worldwide, with deployments up 68% year over year in the first half of 2026. Its fleet logged more than 5.3 million autonomous operating hours in that window, up 24%.
CEO David Pinn's framing is the tell: "Enterprises are not looking for robotics experiments; they need scalable autonomous systems." More than 600 years of continuous autonomous operation in six months. "That is what real-world AI at scale looks like."
Notice what changes at that scale. Buyers stop asking whether one robot can do one task and start asking whether a platform delivers "predictable performance" across "hundreds or thousands of sites." Brain Corp answers with SOC 2 Type II compliance and fleet management, not a livestream.
And notice what these robots are. Floor scrubbers, inventory scanners, wheeled machines in retail aisles and airports. Boring. Single-purpose. Not a five-fingered hand or a backflip anywhere.
The 5.3 million hours are autonomous because there is no operator in Palo Alto looking through the cameras. That is the entire difference, and it is the difference no humanoid deployment has yet demonstrated at any scale.
This is not an argument that humanoids can't get there. Teleoperation data may well train the models that eventually bend the curve. Tacta is reportedly instrumenting gloves with 256 tactile sensors per fingertip precisely because physical-world data is the bottleneck, and it is honest that the data has to come from human hands first.
The argument is narrower and harder to dodge. The metric that proves teleoperation is teaching autonomy already exists, it is auditable, and everyone selling "real work at scale" is choosing not to show it.
Until a humanoid company publishes an intervention rate that falls and a fleet-per-operator ratio that rises, "pilot versus partnership" describes their own product line, not their competitors' clips.
The public markets are already sniffing at the gap. One commenter on Agility's own post noted that Figure is valued at $39 billion and Agility at $2.5 billion — and asked why the one with booked revenue and deployed robots is worth 15x less.
Wrong question, right instinct. The number that should set the valuation isn't revenue or robots deployed. It's how many robots one human can stop watching. Nobody with a humanoid has shown it bending — and the company that has isn't humanoid at all.