Three scenes, the same week, and only one possible reading. In Tokyo, the Humanoids Summit decrees that walking is a closed problem and that the real question is now one of scaling. In Shanghai, the stock exchange approves the public listing of Unitree Robotics, the first embodied-AI player admitted to the A-share market. And between the two, almost inaudible beneath the choreography of the demonstrations, the American NIST acknowledges that, after billions invested, no one yet knows how to measure what these machines can actually do. The industry has listed scale on the market before building the yardstick.
Tokyo shifts the question: from "can it walk?" to "can it scale?"
The sector's doctrine has changed its footing. The central question of the Humanoids Summit Tokyo 2026 is "what has to be true for all of this to genuinely scale" — a clear shift from previous editions, where in 2024 the question was "can it walk?" and then in 2025 "can it work?"[1]. Locomotion — walking, climbing stairs, recovering balance — is now treated there as a solved problem and a baseline expectation for the major platforms; the frontier has moved towards dexterity and fine motor control[2].
The shift is elegant on paper. Yet it presupposes something still to be proven: to declare locomotion "solved" is already to assume that one has an instrument to confirm it. But nothing in the Summit's folklore provides that instrument. The most telling proof sits in a display case: GMO AI & Robotics is showing its G1 robot there, designed to handle airport freight and fitted with a camera-based vision system, whose internal mechanics are in fact supplied by China's Unitree[3]. The robot "at work" at the airport is therefore an assembly with a Japanese badge and Chinese innards — an image of scalability, not a measure of work.
Shanghai capitalises on the narrative: a first listing for embodied AI
On 1 June 2026, the listing review committee of the Shanghai Stock Exchange approved the public listing of Unitree Robotics, making the company the first embodied-AI firm cleared to enter the Chinese A-share market[4]. The operation, carried out on the STAR Market, seals that status of the sector's first listed player[5]. It aims to raise 4.202 billion yuan — around 616 million dollars — by issuing at least 40.45 million new shares[6].
On the accounting side, the trajectory is spectacular. Unitree's revenue rose from 159 million yuan in 2023 to 393 million in 2024, then to 1.699 billion in 2025 — growth of 335% in the last year alone[7]. The volumes follow: deliveries of humanoid robots exceeded 5,500 units in 2025, placing the manufacturer first worldwide by volumes delivered[8]. This is what the public market is invited to capitalise on: a scaling narrative, backed by curves heading upwards.
What the prospectus admits: units sold to labs, a price in freefall
The same filing, however, says something else, in a lower voice. More than 70% of the Unitree humanoids sold in 2025 are destined for research and education — not for a workstation[9]. In other words, what is "scaling" goes mostly to laboratories, not to production lines. And the unit price is not climbing: it is collapsing. The average selling price of Unitree humanoids fell from 593,400 yuan (around 85,000 dollars) in 2023 to 167,600 yuan (around 25,000 dollars) in 2025, a drop of 72% in two years[10].
The company has, moreover, delivered more than 30,000 quadruped robots between 2022 and September 2025[11] — genuine industrial know-how, but one that says nothing about the ability of humanoids to perform measurable work. What is being publicly capitalised, in short, is units shifted to academic buyers and a price in freefall. Not proven work.
The missing yardstick: for the first time since 2015, a neutral arbiter
This is where the fact that the whole dance papers over comes in. According to Aaron Prather, director of the Robotics and Autonomous Systems programme at ASTM International, the NIST's draft benchmark constitutes the first standardised performance benchmark for humanoid robots since the DARPA Robotics Challenge of 2015 — and he stresses that there is still no consensus method for measuring the capabilities of current platforms, despite billions invested[12]. The point is set down in black and white: despite the sums poured into platforms such as Tesla Optimus, Figure, Agility Robotics, Apptronik and Unitree, there existed, until the NIST's draft of April 2026, no consensus method for comparing what these machines can do[13].
The significance of the event lies in the identity of the one speaking. Until now, the standard was written by the manufacturers themselves, judges and parties in their own demonstrations. The NIST is neither a manufacturer, nor a shareholder, nor a broadcaster: it is a neutral third party. And it walks on stage in precisely the week a public market seals the value of an industry that it declares, in the same breath, not yet measurable.
The order of operations, back to front
The simultaneity is no accident of the calendar. The body-counter — Unitree, first by volume, first on the A-share market — wins its public listing on the very day an independent arbiter notes that the work-counter does not yet exist. Tokyo decrees scale, Shanghai capitalises on it, the NIST reminds everyone that we still do not know how to measure it. The demonstrations make for a narrative; they do not make for a metric. And the needle threaded in the display case, like the G1 "working" at the airport, proves nothing as long as no shared instrument comes to attest to it. The industry has had scale listed on the market before inventing the yardstick — it must now hope that the two eventually coincide.