Visitors to Keenon Robotics’ booth during the exhibition season in July and August would have seen mockups of bars, lounges, and theaters. There, the company’s bipedal service robot, XMAN-F1, was often seen popping popcorn, mixing cocktails, and serving chilled drinks. Effectively taking on the role of a human staffer, some might say.
Robots performing real work in these settings isn’t just a stage trick. It reflects the commercial mindset of Keenon Robotics CEO Li Tong. In an interview with 36Kr, one theme Li returned to repeatedly was “commercialization.”
He believes this focus has helped Keenon navigate the volatility that has challenged many commercial robotics firms. It was one of the few to survive the industry’s last major shakeout and remains standing today.
Still, bringing robots into real-world commercial use is easier said than done. According to Li, success requires more than brainstorming in an office. It demands close, on-the-ground observation of business operations to find that elusive overlap between what robots can do and what businesses actually need.
“There’s always a gap between the client’s expectations and your imagination,” he said.
Li recalled Keenon’s early deployment at a hotpot restaurant chain. The company didn’t just send its R&D team, but also product and project managers, to observe and participate. When the robots failed, staff stepped in to deliver food themselves.
To fine-tune its service model, Keenon’s team had to understand every operational detail: how workflows were structured, what each job entailed, how new staff were trained, and what the internal performance standards were.
At that hotpot restaurant, Li noted, there were more than 20 job functions, including both part-time and full-time food runners. Each runner wore a wristband linked to a scanner and earned RMB 0.5 (USD 0.07) per trip by swiping it after each delivery.
So how does one determine whether a robot is commercially viable? Li proposed a concept he called “positionalization.” A robot is ready for large-scale deployment only when it can fully replace a human role and outperform it in efficiency. Anything less, he argued, is just a flashy demo.
Keenon estimates its robots can replace a full-time human position at one-third to one-half the cost. “Robots are labor. We joke that we’re essentially a labor company,” Li said. “Commercialization is about whether robots can truly become labor.”
So far, the company has sold more than 100,000 commercial robots. According to IDC’s 2024 report, Keenon leads the global commercial service robot market with a 23% share.
The robots seen delivering items in Burger King and Hilton properties are likely Keenon-made. Its product lineup also includes cleaning robots, and it has a presence in more than 60 countries and regions, including Japan, South Korea, Europe, and North America.
With a stable business in dedicated-function robots, Keenon began shifting in 2023 toward embodied intelligence. The company’s product strategy is moving from task-specific machines to general-purpose humanoid robots.
Li described a long-term roadmap: first deploying robots in fragmented markets, then gathering large volumes of operational data, and finally using improved models to develop more generalized robotic intelligence.
In a conversation with 36Kr during the World Robot Conference, Li discussed how Keenon has cracked commercialization, its transition from purpose-built to general-purpose robots, and its playbook for international expansion. His insights may offer useful reference points for the next generation of robotics entrepreneurs.
The following transcript has been edited and consolidated for brevity and clarity.
A robotics company is, essentially, a labor company
36Kr: As embodied robots gain traction, how are you feeling about this new wave? Compared with the many newcomers, what is Keenon’s edge?
Li Tong (LT): We’ve been embracing the surge of embodied intelligence, which really only began two or three years ago. Everyone’s basically starting from the same line. But what gives us an edge is how deeply we understand the industry and how seriously we take the product. This lets us bring new robots to market quickly, while others are still showing off tech demos.
Also, we know how to globalize, how to understand different local cultures, and design products to match them.
36Kr: You’ve been in robotics for a decade. How do the past couple of years compare with earlier stages of the industry?
LT: It reminds me of the mass innovation and entrepreneurship wave from years back. In the past decade, we’ve seen cycles of excitement and disappointment every two or three years. We’re used to it.
We’re one of the companies that made it through multiple cycles. Just four or five years ago, service robots were barely present in daily life. Back then, maybe 80–100 companies were trying to break into the field.
What’s happening now looks like history repeating itself. The companies that survive will be the ones focused on genuine commercialization. That was true last time too. Those that didn’t commercialize all failed. Today, only a handful of players have built real scale.
36Kr: Everyone agrees commercialization is critical, but many still fail at it. Why?
LT: Because there’s a gap between imagined deployment and actual deployment. Are you truly understanding your customers, or just assuming you do?
A robot company should be embedded in its clients’ operations, learning their workflows, pain points, and willingness to pay. That’s how you find the overlap between product capability and customer value.
Our robots can replace human staff and cost just one-third to one-half as much. That’s what enables commercialization at scale.
We call it “positionalization.” Only when a robot can fully perform a job more efficiently than a human can it be deployed at scale. Otherwise, it’s just a gimmick.
Robots are labor. We joke that we’re essentially a labor company. Commercialization is about whether robots can truly become labor.
36Kr: Your commercialization approach seems highly scenario-focused. Can you give us an example?
LT: Take that hotpot chain I mentioned as an example. There are 20 job functions, with two types of deliverers: part-time and full-time. Each delivery earns RMB 0.5, tracked with a wristband and scanner.
The labor costs are highly transparent. Even the delivery process is split into three roles: one person plates the food, another delivers it, and a third serves the customer.
36Kr: Keenon is among the few names left from the last robot boom. Is your edge granular control over commercialization?
LT: I wouldn’t say it that boldly, but we believe you can’t commercialize from behind a desk. You have to be in the field, watching how customers work and what their pain points really are.
We’ve spent years inside client environments, studying every detail: workflows, role responsibilities, training processes, standards, everything.
36Kr: How does that approach show up in Keenon’s organizational structure?
LT: All our product managers, R&D staff, and project managers go to client sites. If something goes wrong, they step in to deliver the orders themselves.
Pivoting after selling 100,000 units
36Kr: Before 2023, Keenon focused mainly on hotel service robots. Why the shift to humanoid robots now?
LT: We’ve always debated internally whether to save up for a “big reveal” with a general-purpose robot, or take the incremental route? We chose the latter.
We focus on vertical deployments in niche markets, where we gather data as we go. Right now, the bottleneck in robotics is data. Vertical markets generate shareable, repeatable data. With enough iteration, that leads to general-purpose capability. That’s our commercialization logic.
36Kr: How do specialized and general-purpose robots compare in real-world use?
LT: Most of our early robots were dedicated-use. The upside is that they are highly efficient at one job. The downside is they are not transferable. You need a new design for each new job.
General-purpose robots are the opposite. With data and software iterations, they can be reassigned to different tasks. But they can be less efficient. For instance, a specialized tray robot might carry four trays. A humanoid robot? Two hands, two trays.
So some roles are best suited for general-purpose robots, like waitstaff who multitask. Others, like pure delivery roles, are better handled by specialized robots or robotic arms. You don’t need a humanoid robot for everything.
The future lies in hybrid deployments. Use each kind where it works best to solve labor shortages efficiently.
36Kr: General-purpose robots involve major technical and structural changes. How are you repurposing your past decade of experience?
LT: The core tech is the same: mechanics, motors, control systems, algorithms. All of these are developed in-house. The main difference lies in the embodied AI models. And that’s where everyone is starting from the same line.
There’s no industry consensus yet on which model structure works best. Some, like [Unitree Robotics CEO] Wang Xingxing, are skeptical of visual-language-action (VLA) models. We still believe in them. Maybe he’s right, maybe we are. Until mass deployment happens, it’s hard to say.
Don’t operate in a vacuum
36Kr: One of your advantages seems to be that you’ve got real-world deployment and data. You’re not stuck manually collecting data with joystick controls like many others.
LT: That’s right. But even then, you still need operational data, since spatial awareness is part of VLA architecture. Say you have 100 robots collecting data eight hours a day, that’s still not enough. You only get data at scale through mass deployment. It’s a chicken-and-egg problem: no scale, no data.
36Kr: There’s still a noticeable gap between how robots actually perform and what customers expect. What’s your take?
LT: Take our hotel delivery robots. In China, elevator retrofitting is doable. But overseas, elevators are considered special-purpose equipment, and modifying them is prohibitively expensive.
So we thought, why not give robots arms to push the buttons? Turns out, it’s not so simple.
Many hotels have six elevators. After pressing the button, the robot doesn’t know which door will open. Indicator lights vary: some are red, some are shaped differently. Once inside, the robot needs to verify that the elevator is going to the right floor. If it’s crowded, it has to avoid collisions and still make it out before the doors close.
Robots just don’t understand the real world yet. Something as small as pressing an elevator button is surprisingly difficult and requires high generalization.
36Kr: Can embodied AI help solve this?
LT: Definitely. Last year we launched the world’s first dual-arm embodied service robot, XMAN-W3, with both a “brain” and a “little brain.”
Its “brain” is cloud-based and powered by a multimodal large model that handles long-term task planning and complex scene understanding. Our internal testing showed decent generalization. Of course, real-world deployment is messier. But as training continues, its capabilities will improve.
36Kr: When did you start developing embodied intelligence models?
LT: In 2023.
36Kr: Since you’re strong in hardware, why not partner with a third party to build the robot’s “brain?”
LT: History shows that it rarely works. In a vacuum, without scenarios or data, a “brain” just can’t be built properly.
We once tried buying a navigation module—a relatively basic “brain” for sensing, decision-making, and execution. It never worked because the vendor didn’t understand the business context. That’s why we build our own.
Developing a robot “brain” isn’t about publishing a few papers. We create real-life scenarios, then reverse-engineer the model structure based on what’s needed. Only then do we train and tune it.
Going global is non-negotiable
36Kr: According to IDC, you’re now the global market leader in commercial service robots. What’s your take on going overseas?
LT: For robots, international expansion is inevitable. Robots are labor. Overseas, labor is expensive, so robots are worth more. Our prices abroad are several times higher than in China.
Our primary markets are Japan, South Korea, Europe, and North America. These are places where labor is costly and scarce.
We’re the official robotics partner for the Canadian senior care association. When I asked what kinds of robots they needed, the answer was: “All of them.”
36Kr: How accepting are international consumers of robots?
LT: Japan and South Korea are more receptive. Maybe it’s because people there grew up watching Doraemon, so they naturally see robots as human allies.
The robots we sell in Japan are designed to be cute, making them feel like companions. In the West, though, people grew up with The Terminator. They want robots to look tough.
36Kr: Do you work with local partners to expand abroad?
LT: Yes, local partners are key. In Japan, we work with SoftBank. In South Korea, with KT Corporation. In the Middle East, we’ve partnered with Saudi Aramco. Local backing is crucial for breaking into overseas markets.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.