Staff Reporter |
Editor | Wen Shuqi
At the just-wrapped 2026 World Robot Conference (WRC), NUPIAO literally rolled an entire coffee shop onto its booth floor.
There’s still a human barista behind the counter — the robots handle everything out front: shuttling drinks, clearing tables, and hauling away trash. Unlike the typical roped-off robot demos you see everywhere, visitors could actually walk right into this coffee space, place an order, and sip their brew surrounded by working machines.
Zhang Yufeng, founder and CEO of NUPIAO, is upfront that the robots’ service speed “might be a bit slower than you’d expect” — safety comes first, always. For him, this showcase was really about one question: when a robot drops into an environment where people are constantly moving and every tabletop is a moving target, can it still get the job done, reliably, over and over again?
Throughout WRC, Zhang kept circling back to that exact theme in interviews with NUPIAO and other media outlets.
“Embodied intelligence right now is somewhere between the toddler stage and adolescence,” Zhang says. “It can already do stuff — but complex tasks can’t be mastered overnight.”
For the whole industry, the next battleground isn’t about showing off more flashy moves — it’s about whether robots can hold down a job, hour after hour, in messy real-world settings.
That’s the starkest gap in humanoid robotics today. IDC data shows global humanoid shipments topped 18,000 units in 2025, but over 85% still live in entertainment shows, education research, data collection, and guided tours. Interact Analysis pegs 2025 global production above 20,000 units — yet only about 10% actually work in real operational environments.
The numbers tell a clear story: production is ramping fast, but robots truly embedded in factories and stores remain a rare breed. Interact Analysis concludes the inflection point for large-scale humanoid commercialization may not arrive until after 2032.
NUPIAO’s choice of a coffee shop as its beachhead is no accident — it matches where robot capabilities actually stand today. Zhang explains that coffee shop tasks and SKUs are relatively contained: serving, tidying, cleaning, resetting chairs and tables all have clear boundaries. But the lighting shifts, foot traffic flows, objects move, and customers leave behind personal stuff. Unlike the chaos of a home where there are practically no boundaries on objects or tasks, a coffee shop can train generalization while also being replicable chain after chain.
His shorthand for it: industrial scenes are for polishing skills; commercial scenes are for stress-testing generalization.
On the industrial front, NUPIAO zeroes in on tasks that traditional robotic arms can’t handle well but still carry clear economic value. Zhang cites car seat belts as a prime example. They’re flexible objects — feeding the webbing and buckle into a test rig or fixture looks like a dead-simple motion, but the deformation, force, and position change every single time. Fixed trajectories simply can’t cover it. These jobs still lean heavily on human labor, and NUPIAO sees them as a wedge into automotive parts production lines.
Commercial settings, by contrast, are all about one thing: can the robot keep working when you drop it into a brand-new environment?
Zhang is candid that NUPIAO’s Zero Shot capability is “still a bit short,” while Few Shot performs much better — meaning a new store still needs a small batch of adaptation data. In his view, the imitation-learning-heavy approach demands tons of repetitive data, which is why NUPIAO bet on its MWA technical route built around latent-space world models and reinforcement learning — aiming for higher data efficiency and better cross-scenario generalization.
That thinking traces straight back to Zhang’s past life in autonomous driving.
Zhang spent nearly eight years at Horizon Robotics. Joining in 2017, he rose through vice president, president of the intelligent vehicle business unit, company director, and member of the operating management committee — leading teams through the full arc from R&D to scaled delivery. In the first half of 2025, he left Horizon to found NUPIAO, which launched that same year.
And he didn’t exactly burn bridges. When NUPIAO closed its first 300 million yuan funding round in November 2025, Horizon was among the follow-on investors. Horizon founder and CEO Yu Kai called Zhang “a core comrade-in-arms for many years” and credited him with leading the team through a full-cycle leap.
Ask why he switched lanes, and his answer is refreshingly simple. Zhang tells NUPIAO that autonomous driving hasn’t hit its endgame, but the phase-based landscape has already taken shape — he wanted into an earlier-stage incremental market.
The other reason? He just turned 40 in early 2025. “I wanted to open a new decade of life.”
Capital moved just as fast. Starting with that November 2025 round, NUPIAO has stacked multiple consecutive financings. To date, the company has closed over $200 million in angel funding, with its nearly $200 million Pre-A round close to wrapping up.
For all that funding, NUPIAO is still in early delivery mode. This is the company’s first full delivery year, and Zhang’s target for complete machines is “a few hundred units.” The more pressing question is whether the product can graduate from prototypes and pilots to steady, repeatable delivery.
Zhang has also carried over a hard-won engineering playbook from autonomous driving. That world taught him something unforgettable: going from a working demo to mass production can easily take 100 times the effort of building the demo in the first place.
He argues that once physical AI truly enters mass production, the problems stretch far beyond algorithms — they land on hardware, embedded systems, safety, data loops, supply chains, and long-term reliability.
That’s exactly why NUPIAO isn’t chasing university research labs right now, focusing instead on industrial, commercial, and developer ecosystems.
According to previously disclosed information, NUPIAO has signed global orders totaling 700 million yuan — including a 500 million yuan deal with Envision Group, marking the first 100-million-yuan overseas order in China’s embodied intelligence manipulation sector. The first batch of robots has already shipped to Envision’s battery gigafactory in France.
But orders and deliveries are just step one. As robots actually hit production lines, supply chain maturity and long-term reliability become painfully concrete issues.
Zhang tells NUPIAO that China’s robot supply chain is “very strong” — but strong doesn’t mean mature. Humanoid robots haven’t even developed the kind of standardization consensus the auto industry takes for granted.
He points to upper-body joints as an example. Planetary reducers are cheap and explosive; harmonic reducers offer better precision. Should you compensate with beefier algorithms on cheaper hardware, or splurge on pricier hardware to ease the algorithmic burden? The industry is still figuring that out in real time. Linear modules, meanwhile, have plenty of room to improve on both maturity and cost.
The reliability gap is even starker. He notes that industrial-grade two-finger grippers can handle tens of millions of cycles with a two-to-three-year lifespan, and traditional industrial robots are expected to hit mean time between failures of 10,000+ hours. But plenty of today’s dexterous hands and robot joints don’t come close to that bar.
“China’s embodied industry is incredibly strong on the ‘have vs. have-not’ front, but when it comes to cost and consistency, the whole industry still has a long road ahead,” Zhang says.
That’s the hardest math in commercializing this industry. Whether a robot can replace a job position doesn’t just hinge on the unit price — it’s about success rate, speed, maintenance costs, and how long it keeps running after deployment.
That’s why NUPIAO is prioritizing scenarios where the economics work out more readily. Zhang mentions that labor costs in South Korea are relatively higher — one reason the company is pushing commercial services there. Domestically, the plan is to complete another iteration round and safety validation of its coffee retail product between September and October, then move toward regular operations.
Zhang doesn’t see the coffee shop ending up as a “staffless store,” and the home isn’t a market NUPIAO is rushing into anytime soon. For a company barely a year old, the nearer goal is straightforward: get robots to do real, specific jobs in industrial and commercial settings — and only then worry about generalizing to more complex scenarios.
