Digit's most famous feature is gone

Unitree earned its market leader position by building cheap humanoids. Earlier this year, that strategy stopped buying them the lead.

The old Unitree formula was simple: make humanoids cheap enough that universities, researchers and developers could actually buy them. The H1, G1, R1 and H2 put Unitree well below many Western alternatives, with the G1 currently listed at $13,500, the R1 from $4,900 and the H2 at $29,900. Unitree shipped more than 5,500 humanoids in 2025, ranking first globally by volume according to its IPO materials. Cheap was the strategy, and those shipment figures show that strategy was working.

Then 2026 rolled around and the ranking flipped. The San Francisco research firm Smart Analytics Global reported on August 10 of this year that AGIBOT shipped roughly 8,400 humanoids in the first half of 2026, giving it an estimated 44% of the global market, with shipments up 562% year on year. Unitree shipped about 5,900, or 31%. Those are SAG estimates, not audited industry figures. And AGIBOT did not get there by undercutting Unitree. It got there by selling more kinds of machines: full-size A-series bipeds, smaller X-series units and wheeled G-series robots, and moving into markets Unitree barely touches.

So what, exactly, was Unitree's low price buying? I checked the numbers, and in the same six months AGIBOT took the shipment lead. Unitree's unaudited H1 2026 revenue grew 48.5% to RMB1.15 billion (approx. $172M), but adjusted profit fell 19.34% to RMB244 million (approx. $36.4M), on higher R&D and sales spending. Unitree is still profitable, and its FY2025 numbers are at least audited, which AGIBOT's aren't: no prospectus, no interim report, no audited financials of any kind. AGIBOT's only public number is its own CEO's claim that 2025 revenue 'exceeded RMB1 billion,' with no profit, margin or cash position to check it against. Reuters reported in October 2025 that AGIBOT was targeting a HK$40–50 billion Hong Kong listing; Unitree, meanwhile, is worth roughly 50–53% less than its Shanghai debut peak. Different exchanges, not directly comparable, but the direction says the same thing: Unitree has the receipts, and it's worth less than it was. AGIBOT has a promise, somehow that's worth more.

Two years of robotics marketing taught us to treat low unit cost as the path to humanoid dominance. The first six months of actual competition suggest it may have bought Unitree an installed base instead: thousands of researchers and developers already building around its machines. That is a real asset, and a six-month shipment table cannot measure it. Whether that installed base outlasts AGIBOT's shipment count is the thing to watch over the coming quarters.

Agility’s own announcement of Digit 5 leaves out the most revealing part of the redesign: the legs it is replacing were the point of the research that made Digit possible.

Cassie, the research robot that preceded Digit, was built around dynamic walking and running. Agility traces its origins to more than a decade of legged-locomotion research. Cassie then produced results that made the strange-looking legs hard to dismiss: Agility records a 100-metre world record, while a 2025 research paper notes that Cassie could walk 5km on one battery. The paper describes the backward-facing knee and pantograph structure (a linked mechanism that lets the leg change length) as a design that enabled agile locomotion but looked more like a bird than a person.

That geometry is now being retired because the job changed. Digit is already working in warehouses, and Agility says its robots have logged more than 65,000 hours of operational deployment. On September 15, Agility announced Digit 5 with forward-bending legs and a claimed 50-pound repetitive lifting capacity. The company says the robot can squat fully and sit down when a person comes too close. Its release presents these as engineering upgrades. It discusses payload, new cycloidal actuators and years of customer feedback. It does not say that the backward-bending geometry inherited from Cassie is gone. That fact appears in interviews Agility co-founder and chief robot officer Jonathan Hurst gave instead. In July, Hurst told A3 that Cassie’s ankles hit the ground when it squatted, limiting how far it could get down to lift from the floor.

This is more than a hardware refresh. For roughly fifteen years, legged robotics treated dynamic locomotion as the hard problem worth solving: walk well, run well, stay upright. The commercial job turned out to be different. A warehouse robot needs to get low, lift a box, stand back up and get out of a person’s way.

The field spent years solving how a robot moves, then found that the paying customer mostly wants a robot that can do the human’s job efficiently. Hurst’s description of the result: “Digit's new legs look much more like a person's legs, which, again, goes back to the robot's intended use case... This is where the engineering took us.” Agility didn't set out to make it human-shaped. Making it do the job did that, without being asked.

Two stories this week, one lesson: Unitree got cheaper and still lost the lead. Agility perfected a kind of walking its own customers didn't need. Both companies spent years optimizing for the metric the industry agreed mattered. Neither metric was the one a paying customer was checking.