The key race: humanoid robots move from novelty to factory floor as shipments keep outrunning the AI models meant to justify them
2026 is the year humanoid robotics stopped being defined by a new body and started being defined by whether an operator can collect data, train a policy, and redeploy it repeatedly and economically at scale. Real deployments accumulated through the year rather than staying pilots on paper: Figure AI's units contributed to BMW's Spartanburg, South Carolina line producing more than 30,000 X3 vehicles; Japan Airlines and GMO Internet Group began a Haneda Airport trial in May 2026, running through 2028, using Unitree G1 units for baggage handling, ground-equipment operation and cabin cleaning; and Boston Dynamics' Atlas shipped its first fully committed 2026 customer deployments to Hyundai's Robotics Metaplant Application Center and Google DeepMind. The global robotics market reached roughly $38 billion in 2026, a 34 percent jump from the prior year and the fastest growth the sector has logged in a decade, and analysts now project humanoid shipments alone could exceed 510,000 units by 2030, a compound annual growth rate near 95 percent.
That volume growth keeps arriving faster than the software underneath it. Independent tracking puts global first-half 2026 humanoid shipments above 22,000 units, up roughly 300 percent year over year, with Shanghai's AgiBot holding the top spot at about 43 percent share (roughly 15,000 cumulative units as of mid-2026) and Unitree Robotics second at about 31 percent, the same Chinese one-two that has led every count taken this year. The sector's real technical race sits one layer beneath that hardware, in the vision-language-action foundation models that let a humanoid generalize a skill rather than replay a scripted motion: Nvidia's open-weight Isaac GR00T, Google DeepMind's Gemini Robotics line (demonstrated on Apptronik's Apollo as well as on Atlas), and Physical Intelligence's diffusion-based policies now anchor competing approaches that most hardware makers license or build against rather than developing from scratch. China's shipment lead says more about manufacturing capacity than about whether the robots can do useful work without heavy supervision, the gap this whole sector spent the year trying to close.
Who's ahead: Boston Dynamics turns its Hyundai order into ratified labor peace, Unitree's public valuation keeps unwinding, and Figure bets its future on borrowed compute
Boston Dynamics closed out the underlying labor risk behind its own demand signal. Hyundai's union approved the tentative wage agreement ending the automaker's first full walkout in a decade on August 31, with 61.55 percent of the 31,166 members who voted (19,183 workers) in favor, a deal that raises base salaries 4.1 percent, pays a 400 percent-of-base performance bonus, and addresses retirement age and AI-linked job security after a dispute industry estimates say cost Hyundai more than $1.4 billion in lost sales. That clears the way for Hyundai's roadmap toward 30,000 Atlas units a year by 2028 at its dedicated Metaplant near Savannah, Georgia, alongside the 25,000-unit committed volume the company had already put on the record; Hyundai has said publicly that Atlas will not replace workers even as it finalizes exactly the kind of large-scale deployment plan the strike was fought over.
Tesla's Optimus, by contrast, kept the same slow-and-internal profile it has held all year: production started on Fremont's converted first-generation line in late August, but output still feeds only Tesla's internal 'Optimus Academy' training program, no external unit count has been disclosed, and the Optimus V3 public reveal remains undated even after missing its original Q1 target, with independent reporting now framing 2H 2027 as the realistic start for any external or commercial sales. Figure AI took the opposite bet on where the sector's real constraint lies: on September 3 it signed a strategic partnership with cloud provider Nscale worth $3.5 billion, with intent to exceed $6 billion, covering up to 100,000 Nvidia Vera Rubin GPUs to train and run its Helix foundation model, with first hardware deploying in Barstow, Texas in the second half of 2027 and Nscale taking an equity stake in Figure in return; Figure said its own progress is now 'increasingly constrained by the data and compute required to train Helix.'
China's Unitree turned its manufacturing lead into the sector's first public-market benchmark, and that benchmark kept moving the wrong way: its Shanghai IPO priced August 5-6 at roughly a $9 billion valuation, surged more than 460 percent on its August 19 trading debut to a peak near $66 billion, ahead of Baidu and JD.com, then slid more than 50 percent from that peak by September 2 as first-half results confirmed continued margin pressure (adjusted net profit down 19.34 percent year over year) even as revenue beat prospectus guidance. AgiBot's own Hong Kong IPO process, opened July 24 targeting a $5.1-6.4 billion valuation, and Oregon-based Agility Robotics' $2.5 billion SPAC merger with Churchill Capital Corp XI both remain unpriced and awaiting shareholder and regulatory approval as of this update, effectively unchanged from where both stood a month ago.
What decides it: whether the AI models catch up to the hardware, and whether the robots are safe enough to trust once they leave the demo floor
HSBC's own read on Unitree's IPO frenzy names the central risk directly: the bank's analysts wrote that 'in the absence of a significant improvement in the AI model capability of robot makers, the current humanoid robot shipment upcycle is unlikely to be sustained over the next 1-2 years,' a warning sharpened by Unitree's own first-half guidance implying revenue growth slowing to 36-45 percent (from 333 percent a year earlier) against a valuation multiple that peaked north of 200 times trailing earnings and leaves no room for that deceleration to continue. That is the sector's core tension in one company: shipment volume is real and growing, but the AI capability needed to make each unit worth its price tag is advancing on a slower, less certain curve.
A second, newer risk surfaced in August when security researcher Olivier Laflamme disclosed two root remote-code-execution flaws in Unitree's G1 EDU robot, CVE-2026-76639 and CVE-2026-76640, the second exploitable over Bluetooth without prior pairing and flagged as potentially wormable, meaning a single compromised G1 could in theory help spread the same exploit to other vulnerable units in radio range; Unitree patched the flaws within about two months of disclosure. The timing mattered: G1 units had already moved from research labs into higher-stakes settings, tarmac baggage handling at Haneda Airport since May and a hostage-negotiation and standoff-response role at a North Carolina sheriff's office since July, adding a fleet-scale cybersecurity dimension to a national-security scrutiny narrative that had previously focused on the Pentagon's Section 1260H list and the FCC's Covered List addition. Labor remains the sector's third open variable, and Hyundai's own ratified settlement, a dispute over exactly the AI-and-robots job-security question every other manufacturer pushing humanoids onto its own floors will eventually face, is the clearest evidence yet that a rollout this size cannot proceed on engineering timelines alone.
The money and the rules: Beijing makes embodied AI a national priority as global humanoid funding blows past its own record
China's 15th Five-Year Plan (2026-2030) elevated embodied intelligence to its own distinct strategic industrial priority for the first time, alongside quantum technology and 6G, backed by a national venture fund targeting roughly $138 billion for robotics, AI and adjacent innovation and more than $26 billion in local funds assembled by Beijing, Shenzhen and other cities since late 2024, some offering free land, subsidized rent, and buyer rebates worth about 10 percent of a robot's sale price. The plan sets an explicit national target of 100,000 humanoid robots deployed by 2027 and calls for a complete domestic supply chain covering actuators, dexterous hands and perception systems, the clearest sign yet that China is treating humanoid manufacturing as state industrial policy rather than leaving it purely to its own venture market.
Private capital kept pace globally: humanoid-specific startup funding hit roughly $8.6 billion in 2026, already 1.8 times all of 2025, inside a broader robotics venture total north of $23 billion for the year. Figure AI remains the best-funded pure-play humanoid company at roughly $39 billion following the Nscale deal on about $2.34 billion raised; Apptronik closed a $520 million Series A extension in February at a $5.3 billion valuation (nearly $1 billion raised total) backed by Google, Mercedes-Benz and the Qatar Investment Authority; and Germany's untracked Neura Robotics raised roughly $1.4 billion in June from Tether, Nvidia, Amazon, Bosch and Schaeffler at a valuation reported between $7 billion and $12 billion, funding a stated goal of 5 million humanoid robots by 2030. Unitree's own valuation arc, from a $9 billion IPO price to a $66 billion debut peak to a slide of more than half by early September, is now the sector's most visible live test of whether that capital is pricing a real production ramp or a speculative one.
What to watch through 2027-2028
Watch whether Tesla's Optimus V3 reaches a public reveal date and disclosed external production numbers, and whether Fremont clears even a fraction of an original 50,000-unit 2026 target before the 'flat and long' ramp Musk has promised turns upward. Watch Hyundai's Atlas rollout hold its 30,000-unit-a-year-by-2028 Metaplant target now that the wage deal addressing AI-linked job security is ratified, and watch Figure's Nscale-funded compute come online in Barstow, Texas in the second half of 2027 and whether it visibly narrows the gap between Helix and the shipment volumes Chinese makers already claim. Watch AgiBot and Agility Robotics finally convert their respective Hong Kong IPO process and SPAC merger into priced, trading listings after months in registration.
Watch whether China's 100,000-humanoid-by-2027 national target converts local-government subsidy money into robots doing real work rather than robots merely shipped, and watch Unitree's stock for a floor, or a further leg down, as the clearest public read on whether the sector's shipment upcycle is durable. Watch for patch-adoption follow-through on Unitree G1's wormable Bluetooth flaw as more fleets deploy into critical infrastructure and public-safety roles, and watch whether Neura Robotics, Apptronik, or another well-funded challenger converts 2026's record venture money into a foundation-model compute bet of a scale comparable to Figure's Nscale deal.
Sources
- IDC - What the 2026 Humanoid Robot Half Marathon reveals about industry progress
- Nscale - Nscale and Figure sign strategic partnership to power the next generation of physical AI
- The Korea Herald - Hyundai Motor union approves wage deal, ending months of labor strife
- The Hacker News - Two Unitree G1 EDU humanoid robot flaws enable root RCE, one starts over Bluetooth
- The Diplomat - China's new five-year plan prioritizes robotics
- Bloomberg - Unitree plunges 50% from peak in fast reversal after huge debut pop
