The humanoid robotics industry has spent three years arguing about legs. How many actuators per leg, how much torque per knee, how gracefully a machine can descend a stage after a dance routine. Then last week the largest buyer ever to enter the market filed its order, and the order is a quiet rebuttal of the entire discussion. Toyota Motor group, per Nikkei Asia, plans to deploy 400,000 in-house-developed robots across its worldwide production sites starting in 2028, backed by factory-upgrade investment that could reach ¥1 trillion, about $6.4 billion, every year. The flagship machine carrying that program is ELEY: a 50-kilogram humanoid that rolls on wheels, has two-fingered hands, and was engineered around a research agenda its creators describe as gentle touch.
Scale first, because the scale is the story. Of the 400,000 units, 150,000 would go into Toyota’s own automotive plants and another 250,000 into group company facilities producing components and materials, spanning roughly 60 factories globally, Automotive News reports. For context, the entire rest of the humanoid industry is fighting over production targets in the low thousands. Tesla built a few hundred Optimus units in 2025 against a 5,000-unit goal. One buyer has now committed to a fleet two orders of magnitude larger than the global humanoid installed base, on a timeline that starts in roughly 15 months.
What the money actually says
It is worth pausing on what ¥1 trillion per year buys, because the framing matters. The investment figure covers factory upgrades writ large, not robot purchase price alone: retooling lines, reworking logistics, integrating fleets. But the direction of the capital is unambiguous. Toyota already operates one of the most automated production systems on earth, and it is choosing to pour additional billions annually into general-purpose machines that work alongside people rather than into more fixed automation. The International Federation of Robotics, cited by Ars Technica, counts China’s operational industrial robot stock above two million units with Japan second at roughly 450,500. Toyota’s plan alone would add a fleet approaching a fifth of Japan’s entire current installed base, and it would do so with humanoids rather than caged arms.
That choice is the most expensive vote of confidence the field has received. Venture rounds validate founders. A ¥1-trillion-a-year capital program from the world’s largest automaker validates the category, and it does so with a buyer’s discipline: Toyota is not investing in humanoid startups, it is deploying robots it built itself, for work it has watched its own factories perform for decades.
ELEY: the deliberately boring humanoid
ELEY, which stands for Embodied Learning robot for Enhanced Yield, looks like what it is: a tool that won an internal engineering argument. It is a wheeled platform powered by battery or cord, with two-fingered hands rather than five-fingered anthropomorphic grippers, and it weighs 50 kilograms. Every one of those decisions sacrifices demo theatrics for reliability, and Toyota has been transparent about the trade. Where Tesla’s Optimus, Figure’s models and Hyundai’s Atlas pursue bipedal mobility on the theory that the world is built for human bodies, Toyota’s Frontier Research Center concluded that its factories have smooth floors, and that a machine that cannot fall down cannot fall over.
The design lineage is documented on Toyota’s own research pages, and it is older and more substantive than most of the humanoid field combined. ELEY descends from the Human Support Robot unveiled in 2012, through the Welwalk rehabilitation robot and the teleoperated T-HR3, with intermediate technologies already productized in the Potaro hospital transport robot and the KumiPro factory robot running at Toyota Motor East Japan today. This is not an automaker chasing a trend. It is a 14-year robotics program reaching its factory deployment phase.
The technical core of ELEY is an actuator story, and it inverts the industry’s usual spec-sheet logic. Toyota’s researchers found that the robots most likely to break were the ones that touched things: a slight grasp deviation sends unexpected forces through an arm geared for torque, causing damage. Their answer was the quasi-direct-drive actuator, pairing a high-torque motor with a gear reduction ratio of 10:1 or less. Low gearing sacrifices the compact torque density that conventional humanoid actuators advertise, but it buys backdrivability: the joint yields when pushed, so contact becomes compliant instead of catastrophic. In an assembly task, where a robot will graze fixtures and bump workpieces thousands of times a shift, the ability to give way quietly is worth more than peak torque.
Even ELEY’s silhouette encodes the philosophy. Its joint structure and proportions were modeled on the average adult Japanese male, with joint spacing and arm thickness matched to human norms, and its designers added a scapular axis, shoulder blades, because humans lean on their shoulder mechanics for two-handed pushing and reaching. The robot imitates the human body where imitation helps manipulation, and abandons it, legs first, where it does not.
Takumi capture: the finger jigs
The training pipeline is stranger and more interesting than the hardware. ELEY learns from demonstration, and the demonstrators are Toyota’s own master craftsmen, the takumi. Roughly 18,000 veteran workers across the company’s 60 factories hold skills that took decades to build, and most of them are approaching retirement in a labor market that cannot train replacements fast enough to replace them. The ELEY program is, at bottom, a race to convert tacit human expertise into machine behavior before it walks out the door.
The capture method is almost comically physical. Workers wear jigs shaped like the robot’s own two-fingered hands over their fingers, then perform assembly tasks normally. The jigs force human motion into the robot’s motor vocabulary, so every demonstration produces directly transferable training data rather than motion that must be re-targeted onto a dissimilar body. At a briefing at Toyota’s European headquarters, the system reached near-perfect accuracy folding T-shirts after 1,500 practice sessions over two weeks, as M4SNews details. Underneath sits Toyota Research Institute’s Large Behavior Model work, diffusion-based generative policies that learn physical skills from small numbers of human demonstrations rather than millions of simulated trials.
The strategic implication deserves emphasis: Toyota does not have a data problem. Every humanoid startup on earth is burning capital on teleoperation fleets to scrape manipulation data from rented warehouses. Toyota has 60 factories generating authentic production tasks, 18,000 experts to demonstrate them, and a captive customer with a 400,000-unit order book. The data flywheel the industry keeps promising investors already exists inside Toyota’s walls, and it is powered by people wearing robot fingers.
2028: the wheeled and the legged meet on the same floors
Toyota’s program lands in the same year as its most direct rival’s. Hyundai plans to begin deploying Boston Dynamics’ Atlas at its plants starting in 2028, targeting up to 25,000 units across Hyundai and Kia facilities. On September 21, Boston Dynamics opened a robotics training center at Hyundai Motor Group Metaplant America in Georgia, focused on teaching Atlas real manufacturing sequences such as parts sequencing for assembly. The automotive industry has thus quietly staged the controlled experiment the humanoid field has debated since 2023: two manufacturing giants, the same start year, one wheeled and grip-simplified, one bipedal and fully anthropomorphic.
Toyota is hedging its own bet. The company also partners with Agility Robotics and runs bipedal Digit robots at sites including Toyota Motor Manufacturing Canada. That makes Toyota the first customer in history to operate wheeled humanoids, legged humanoids and legacy automation under one production umbrella, which means its internal utilization data will eventually settle the form-factor argument with a rigor no demo ever can.
The labor answer nobody quite believes
Executive Vice President Hiroki Nakajima framed the program to Nikkei as aiming for “a world where robots coexist with humans, rather than replacing them.” The arithmetic makes the framing necessary and the skepticism understandable. Toyota’s Japan-facing workforce is aging out of skilled roles faster than it can be replaced, and the company explicitly plans for robots to help train new human employees, not only learn from veterans. But 400,000 machines performing takumi-grade work is a substitution of labor content under any accounting, and the labor movement is watching: when GM installed robots at its flagship EV factory months after laying off 1,300 workers, the backlash wrote itself.
The honest version of Toyota’s position is demographic rather than altruistic. Japan’s working-age population is contracting; the takumi cohort is retiring; the choice on offer is not robot versus human but robot versus lost capability. That argument is true as far as it goes, and it will be tested the first time a human line worker and an ELEY compete for the same task assignment.
What the order settles and what it does not
Three conclusions follow. First, the demand side of the humanoid market now has an anchor tenant. Suppliers of actuators, reducers, batteries and compute can underwrite capacity expansion against a committed buyer at a scale no startup order book provides, and the pull will be felt across the whole component chain. Second, the bipedal consensus is officially broken at the top of the market. The largest deployment commitment in the field’s history went to a wheeled machine with two fingers, on the reasoning that factories reward uptime over anatomy. Legs remain a hypothesis with impressive demos; wheels are a hypothesis with a purchase order. Third, the real moat in humanoid robotics is looking less like model architecture and more like proprietary demonstration data from real work, which favors operators of physical industry over AI labs. Toyota is not trying to win the foundation model race. It is trying to own the training corpus, and it may have just bought the largest one ever assembled.
None of this guarantees execution. The 2028 line in Toyota’s plan will collide with the same realities that humbled every previous humanoid schedule: integration cost, exception rates, safety certification, and the long tail of edge cases that separates a folding demo from a shift on a line. But when the history of this industry’s commercial era is written, the week the world’s largest automaker ordered 400,000 robots, and specified wheels, will mark the moment the humanoid business stopped being a showcase and started being procurement.