Las Vegas, first week of January 2026. Thirty-eight humanoid robots overrun CES — including twenty-one Chinese manufacturers, or 55% of the category. The footage that went round the networks shows two Unitree G1s boxing in a BattleBots arena, the H2 stringing together backflips and kicks, the G1 dancing in synchronised choreography, the R1 in kung-fu poses. Spectacular folklore, many concluded. That diagnosis falls short — and probably runs counter to what actually played out that week.

The teleoperated spectacle, the new sales funnel

What the media coverage largely passed over in silence: the Unitree H2 includes a teleoperation capability via wearable rigs and the Apple Vision Pro. This is not a spec-sheet footnote. Unitree's public xr_teleoperate repository explicitly describes the tool as a data collection toolkit for embodied intelligence, designed to record the human operator's movements — Vision Pro, Quest or PICO — in order to feed imitation learning. The spectacular gesture and the data capture are one and the same gesture.

Analysts present at CES confirmed that the vast majority of the systems on show sat at autonomy level 0 or 1 — scripted or teleoperated. Boston Dynamics' Atlas demonstration at the show was likewise remotely piloted. That fact, rather than invalidating the robots, reveals the model's true architecture: teleoperation is no longer a weakness to be concealed, it is the data-collection mechanism the sector has industrialised.

The rock-bottom price as a data acquisition cost

The Unitree R1 starts at 4,900 USD, in the R1 Air version. It is positioned as a developer-first platform. Reading that price as competitive dumping would be a framing error. Putting a teleoperated body into the hands of thousands of labs, integrators and early adopters at that price means distributing collection sensors en masse: every demo, every teleoperation session feeds the imitation learning dataset. The 4,900 USD R1 is not being sold at a loss to kill off the hardware competition — it is being sold to accelerate the accumulation of motion data.

Unitree, for that matter, announced at CES 2026 an explicit pivot towards a Robot-as-a-Service model, complete with a skills marketplace. The intent is clear: the robot is the terminal; the value lies in the data and skills layer that accumulates above it.

NEURA, LG, AgiBot: three data taps

The same move can be read across the show's other major players, each by its own route.

NEURA Robotics presented the Neuraverse: an "invisible" operating system that connects networked robots and devices to share collective learning. The built-in marketplace lets developers publish, share and monetise robotic skills. NEURA's robots are trained in NEURA Gyms — real, physical environments, not simulation alone — to generate physical-AI data. The training infrastructure is as much the product as the robot.

LG Electronics unveiled the CLOiD, a domestic robot whose VLA (Vision Language Action) model was trained on tens of thousands of hours of housework task data. At the same time, LG took stakes in Figure AI (United States), AgiBot (China) and Dyna Robotics. This investment portfolio is not geographic hedging in the classic financial sense. AgiBot is known precisely for its data learning farms — infrastructure for collecting and training on motion data. LG is not buying diversified hardware bets: it is securing data taps on two continents.

What was autonomous, and went unfilmed

The only genuinely demonstrated autonomy at CES 2026 went almost unnoticed, precisely because it is boring to watch. X-Humanoid (Beijing Humanoid Robot Innovation Center) showed two machines: the Tien Kung 2.0, which sorts industrial parts in full autonomy, and the Tien Kung Ultra, which ran a 21.0975 km half-marathon with no remote control in 2h40m42s. No choreography, no arena. Neither of these feats went viral.

The practical demonstrations were there all the same: making drinks, folding laundry, dealing cards, sorting parts. But it was the backflips and the boxing bouts that shaped the coverage. The hierarchy of attention is itself telling: the spectacular teleoperated robot catches the eye and the share; the useful autonomous one is invisible. The sector has absorbed that filter and turned it to its advantage.

The EngineAI CEO's signal

An incident on the fringes of CES illustrates the latent tension. Zhao Tongyang, founder and CEO of EngineAI, posted a video showing the T800 kicking him, filmed from several angles in a studio, to answer doubts about the authenticity of his robot's capabilities. That an executive should have to supply multi-angle proof to convince people that his robot moves on its own says something about the state of trust in the market — and about the mounting pressure to tell what is autonomous from what is staged.

The shift, not the revolution

In December 2025, this chain was already showing that the choreography is a product spec, that the R1 works as a data probe, and that manufacturers were having to prove on video that a robot moves unaided. CES 2026 is the next step in that same arc: the sector is no longer ashamed of teleoperation, it has turned it into a business model. The dance is no longer a lie to be hidden — it is the customer acquisition cost of a data platform that Unitree, NEURA, LG and AgiBot are now openly fighting over.

Tracking the number of robots at a show, or comparing their acrobatics, is looking at the terminal. The battle is being fought on the layer above: who accumulates the richest motion data, who controls the skills marketplace, who owns the training infrastructure. Las Vegas was the set. The product was the harvest.