Reporter |
Editor | Wen Shuqi
On August 28, AI robotics company Sharpa revealed that it has raised over 4.5 billion RMB in cumulative funding , with backers including industrial capital from Alibaba, Meituan, Tencent, JD.com, and Transsion, as well as institutions like Sequoia Capital China, Qiming Venture Partners, Meituan Longzhu, and Guanghe Ventures.
The same day, a robot-powered restaurant co-developed by Sharpa and DQ made its debut in Shanghai. According to Sharpa, the project runs on a “zero-renovation, fully autonomous, year-round” model, with robots handling the entire Blizzard ice cream-making process inside a normally operating store.
In an interview with NUPIAO and other media outlets, Sharpa co-founder Li Yifan said the key to moving robots from demo to commercialization isn’t just executing a single action—it’s whether you can “fully work through a scenario.” The system has to run end-to-end, customers need to keep using it, and there has to be room for scaling up and cutting costs.
He calls this minimum bar the “threshold of viability.”
In Li’s view, a demo that scores 70, 80, or even 90 points can still showcase technical progress. But once you enter a real-world scenario, the yardstick changes. Only when the entire system’s efficiency and reliability reach a level where customers are willing to keep using it can a scenario be considered truly solved. If you haven’t crossed that line, even outstanding individual capabilities are hard to replicate and scale.
In this particular deployment, the store’s existing kitchen equipment, ingredients, tools, and workflows weren’t redesigned around the robot. The robot has to independently complete all 55 steps of making an Oreo Blizzard—from picking up the cup, adding ingredients, and mixing, to handing it off. A reporter on site timed it: about 6 minutes per cup, while an experienced human employee typically takes two to three minutes.
That also highlights the most pressing issue in robot commercialization right now: robots can do the job, but they’re still a ways off from meeting the efficiency and cost demands of real commercial use.
When asked about return on investment per store, Li told NUPIAO that this single unit isn’t profitable yet. Rather than focusing on immediate returns, he’s more concerned about whether a few long-term variables can keep improving—speed, equipment lifespan, and hardware costs.
Sharpa was founded in late 2024 by Li Yifan, who also serves as CEO of lidar supplier Hesai Technology, along with co-founder and CTO Xiang Shaoqing and chief scientist Sun Kai. The company develops general-purpose robots and core components. Sharpa operates independently from Hesai, with no equity or managerial control ties between them.
Its first product is the SharpaWave, a tactile dexterous hand with 22 active degrees of freedom, which entered mass production in October 2025. Sharpa’s current lineup also includes North, a wheeled whole-body humanoid robot, and CraftNet, an AI model designed for dexterous manipulation.
Rather than focusing on product and tech showcases, Li spent more time in interviews with NUPIAO and other outlets discussing what actually happens when robots enter real scenarios.
In his view, the biggest difference between a demo and a real-world scenario is that the latter doesn’t accept “mostly getting the job done.” Take the ice cream process—55 steps. If a robot can only nail 54 of them long-term, and the last step still needs a human, then you can’t truly pull a person off that position. For a business, the value of such a deployment takes a big hit.
That’s also why Sharpa insists on not modifying the environment—letting robots adapt to real scenarios as they are.
Li believes that if a robot only works by retrofitting stores or customizing tools, it’s essentially closer to traditional automation. Switch to a different store or a different industry, and you’d have to re-adapt everything from scratch. What a general-purpose robot really needs to solve is being able to get the job done in environments originally designed for humans.
However, taking this path also means it’s hard to prioritize speed right now. Li told NUPIAO that there’s a direct trade-off between robot movement speed and reliability. The same workflow can be compressed further, but the probability of errors goes up accordingly. So Sharpa is currently choosing to ensure stability first, then speed things up through training and engineering optimization.
If robots are to scale from one store to a much larger deployment, pricing and long-term reliability are also issues that need solving. Li says the current system is priced in the hundreds of thousands of RMB, with a long-term goal of bringing it down to around 100,000 RMB. On reliability, he believes robots should eventually reach a level comparable to automotive products.
Another unavoidable issue is data. Li argues that compared to simply building data farms or recording human actions via first-person video, data generated from real operations is far more valuable.
In his view, first-person video can capture human movements, but it’s hard to extract force, tactile feedback, and other information during the process. Meanwhile, failure data generated by robots during real work is also a crucial source for model training.
Real-world scenarios also throw up countless variables that are tough to anticipate in simulation. Li gave an example: the different temperature states of an ice cream machine after startup change the viscosity of the ingredients, which in turn affects the robot’s grasping and mixing. These kinds of issues often only surface during actual operation, and then need to be solved by adding training data from various states.
On whether scenario operators are willing to open up their data, Li thinks the order should be reversed: robot companies first need to prove their product is useful in a given scenario, and only then can they potentially obtain data without harming the other party’s commercial interests or privacy. If the product doesn’t solve a problem and the company demands data upfront, the partnership won’t work.
This also shapes Sharpa’s judgment on which application scenarios to pursue. Li says that compared to display-type uses like explaining or performing, he’s more interested in whether robots can genuinely take over human work.
In his view, factories offer tremendous value, but they’re also harder. Highly repetitive work has mostly been automated by traditional equipment already; what’s left tends to be complex, non-repetitive tasks. For humanoid robots to enter these roles, they not only need to “know how to do the job,” but also meet requirements for cycle time, reliability, and cost simultaneously.
From that perspective, this DQ store is more like a long-term commercial validation than proof that robots have already cracked commercialization.
Whether a robot can go from a one-off demo to a replicable product ultimately depends on whether it can work reliably in a real operating environment over the long haul, reduce human intervention, and bring efficiency and cost to a level customers can accept. Only then can a single scenario validation be replicated across more stores. That’s what Li means by crossing the “threshold of viability.”
