The founding wager of the humanoid robotics industry is easy to state and hard to defend: build one machine in the shape of a person, and the whole built world becomes addressable. Stairs, doorways, tools, workbenches, and vehicles were all designed around the human body, so the argument goes, and a body that matches can inherit all of it. That single assumption justifies billions in venture capital, because a robot that can in principle do any task can in principle be sold to everyone.
The summer of 2026 has produced the most interesting counter-evidence yet, and it comes not from skeptics but from two of the most credentialed teams in the field. Walden Robotics, the Toyota Research Institute spinout led by MIT professor Russ Tedrake, emerged from stealth in July with $300 million at a $1.1 billion valuation and a robot that skips legs entirely. Persona AI, founded by former NASA Valkyrie program lead Nicolaus Radford and former Figure CTO Jerry Pratt, has committed to legs, but only because its chosen job, shipyard welding, physically demands them. Both companies started from the same place: find work where a humanoid makes money now. They arrived at opposite bodies. That divergence is the story, because it marks the moment the industry’s one-body assumption stopped being orthodoxy and became a choice among others.
Walden: The Case Against Legs
Walden’s launch, covered in detail by IEEE Spectrum, is the highest-capitalized bet yet that the humanoid’s most iconic feature is also its most dispensable. The company closed a $300 million seed round co-led by Toyota (through Toyota Motor Corp, Toyota Invention Partners, and Toyota Ventures) and Deviation Capital, with participation from NVIDIA, Boeing, Samsung Ventures, Prologis Ventures, and CoreWeave Ventures. The investor list reads like a customer list: an automaker, an aerospace prime, a logistics real-estate giant, a cloud GPU provider. These are companies that buy factory labor at scale, and they funded a robot designed around their own floors.
Tedrake’s argument against legs is disarmingly practical for a man who spent two decades studying them. “I thought about legs for 20 years; that’s the class I teach at MIT,” he told Spectrum. “There are many reasons to build a robot with legs. But the question is, what’s the addressable market? And what percentage of it is covered by a wheeled base?” The honest answer to the second question, for manufacturing and logistics, is most of it. Factory floors are flat. The work is at waist height. The robot needs to move between stations, not climb stairs.
The wheeled base buys three things that matter more than romance. First, batteries: a heavy base near the floor adds stability rather than risk, so Walden can fill it with cells and delete the mid-shift recharge problem that haunts bipedal fleets. Second, static stability: a robot on wheels does not fall over, which collapses the safety case that currently keeps legged humanoids fenced off from human coworkers. “Factories already have autonomous mobile wheeled robots,” Tedrake notes. “They already have safety cases built around AMRs. You can piggyback on that with a wheeled base.” Third, payload: the chonky base carries the mass needed for useful manufacturing work.
The same pragmatism governs the hands. While the frontier chases five-fingered anthropomorphic hands, Walden built rugged grippers, and for a specific reason: “We have been deployed in a Toyota factory, and at the end of the week, the hands take a beating, so we built hands that can take that,” Tedrake says. “I have not seen a more dexterous hand that could have done the work our hand has done.” Durability, not dexterity, is the binding constraint in production.
What makes Walden more than a clever wheeled-AMR company is its software lineage. The team behind it pioneered Diffusion Policy at TRI, and the company’s robots run on Large Behavior Models, the frontier model class that lets them acquire new tasks by building on previously learned skills. That heritage shows up in deployment speed: Walden launched out of TRI in January 2026, entered a Toyota plant in North America in February, and moved from first pilot to real production work in under two months. Toyota’s CTO Hiroki Nakajima framed the partnership in explicitly Toyota terms, invoking kaizen and jidoka and robots that “provide value from day one in real-world work environments.” For an automaker that spent a decade being criticized as late to robotics, buying the people-first framing wholesale is a deliberate positioning move.
Persona: The Case Where Legs Earn Their Keep
Persona AI’s path to the opposite conclusion runs through the same economic filter. When Spectrum first profiled the company in 2024, it was, in Radford’s words, “all over the place”: warehousing, automotive, even the home. Two years of sharpening produced a radically narrower answer: welding, in shipyards, with human tools, for the world’s largest shipbuilder.
The physics of the choice are elegant. Welding is a skilled trade in chronic shortage, and it is taxing, high-precision physical labor. Yet the task itself is unusually robot-friendly. “Because you can only weld as fast as metal melts, the top speed for the task is an easily manageable centimeter per second,” Pratt explains. Quality welding means millimeter-scale repeated motions sustained over hours, exactly where robots outperform tiring, bored humans. And the volume is staggering: a single ship requires hundreds of kilometers of linear welds. Persona’s two public partners, HD Hyundai and steel giant POSCO, are both labor-constrained and running significant backlogs, which inverts the usual robotics pitch. “We want to help our customers’ top line, not necessarily their bottom line,” Radford says. “Even if our robot was more expensive than a human, that would still be valuable to these companies, because it could unlock additional revenue.” A robot that costs more than a welder but adds a welder who does not exist is not a cost-cutting tool. It is capacity.
Notably, Persona explicitly rejected the sector’s default targeting of cheap, unskilled labor. “We started forming this thesis around skilled trades and tool usage,” Radford says, calling the approach “last-mover advantage”: watch everything everyone else tries, then pick differently. Skilled trades are expensive enough that the robot’s economics clear the bar without heroic cost reduction, and the customer base consolidates: Radford says Persona wants customers who can absorb hundreds of robots per location, not ones or twos.
And here legs win, for once, on merit rather than ideology. “These open-air shipyards are a couple hundred meters long, with horizontal and vertical spars that you have to step over all the time,” Radford says. “You’ve got to work on the ground, overhead, and through portholes.” A wheeled base that cannot cross keel blocks or climb into a hull section is worthless at that worksite, no matter how stable it is. Persona even considered four or more legs before settling on two as the least disruptive to existing shipyard rhythm. The safety story, normally bipedalism’s weak point, is also softened: shipyard workers are already trained around heavy industrial hazards, and the environment tolerates a robot that a suburban kitchen would not.
The Convergence Beneath the Divergence
Set the two robots side by side and the differences look fundamental: wheels versus legs, grippers versus tool use, flat factory versus open-air yard. The similarities are more consequential.
Both companies chose environments where the safety case is pre-built by context, AMR conventions on Walden’s factory floors, industrial hazard training in Persona’s yards. Both target utilization rather than novelty: Tedrake demands applications “where the robot is used 24 hours a day, 7 days a week,” while Persona’s backlog economics effectively guarantee full utilization from day one. Both design for durability over demos. Both defer the home indefinitely: Pratt notes that consumer robotics would impose “a lot of competition and a ton of price pressure” that industrial niches avoid. And both describe their long-term ambition in deliberately modest language. Tedrake prefers “multipurpose” over “general-purpose” because it “moderates expectations in the near term.” Persona’s expansive goal, to become “the largest repository of industrial skills,” is a data strategy disguised as a hardware business: tens of thousands of hours of expert demonstration, collected weld by weld, is the asset that compounds.
This is what an industry-wide sobering looks like when it reaches hardware design. Spectrum’s own framing of the Walden piece observed that the sector has moved from a “mad rush” toward pointed questions about practicality and value. The pattern across both interviews is identical: start from the wage rate and the backlog, derive the environment, and let the environment pick the body. The body is an output, not an input.
What It Means for the Bipedal Majors
The uncomfortable implication lands on the companies whose valuations are anchored to the one-body wager. Figure, Tesla Optimus, Apptronik, and Agility are all selling horizontal generality: one platform, every warehouse, eventually every home. That thesis is not wrong so much as unfalsifiable for now, and unfalsifiable theses are expensive to hold. Walden’s $1.1 billion valuation is explicitly not a bet on the human form; it is a bet on Diffusion Policy-lineage software and a Toyota-shaped customer pipeline. Persona’s investors bought a welding company with a skill-library option attached.
The pressure this creates is directional. If wheeled platforms capture the flat-floor manufacturing and logistics market, the highest-volume near-term deployments, the bipedal majors are left competing for the environments where legs are truly required: construction sites, stairs, the eventual home. Those are harder, later, and in the home’s case, price-constrained. The middle ground may go to whoever acknowledges fastest that “humanoid” is a marketing category, not an engineering requirement.
None of this settles the form-factor question. Persona’s own trajectory shows how contingent the answers are: legs in shipyards today, and its founders point to grinding, painting, and fabrication as adjacent expansions. Walden’s LBMs are form-agnostic in principle, and nothing stops the software from eventually walking. The lesson of August 2026 is not that wheels beat legs or the reverse. It is that the question now has a defensible answer on both sides, and every serious operator, investor, and OEM should be more suspicious of any robot whose body was chosen before its job.
Sources: Walden Robotics launch announcement (July 15, 2026); IEEE Spectrum, “Why Walden’s Humanoid Robots Run on Wheels, Not Hype” (Aug 3, 2026); IEEE Spectrum, “How Persona AI Makes Humanoids Pay Off In Welding” (Aug 24, 2026).