Hotels may be the first stop for humanoid robots, says Unix AI founder

Since the end of the World Humanoid Robot Games, Unix AI’s customer hotline has been ringing nonstop. “In the second week after the competition, more than a dozen hotel clients came to visit us,” said Yang Fengyu, founder and CEO of Unix AI.

Earlier in August, Unix AI won two gold medals and one silver in the hotel cleaning and guest reception categories of the international event. That success quickly drew the attention of hotels, retirement homes, and other operators exploring service robotics.

Both competitions tested robots on generalization, dexterity, and speed. The cleaning task required collecting scattered items in a room as quickly as possible. The reception task involved taking a guest’s suitcase and delivering it to a designated endpoint.

Unix AI’s strong performance reflected accumulated experience. Its robots were already deployed in “quasi-consumer” cleaning scenarios, particularly in hotels, where they learned on the job while gathering data. In hotel rooms, they clean, tidy, and remove trash. While still slower than human staff, the relatively high tolerance for error in these settings allows them to work behind closed doors at their own pace, Yang said.

He believes the skills honed in such environments will eventually transfer to households, restaurants, cafes, and even fast-food outlets. The company is already in the early stages of small-scale deliveries, with contracts signed with hotel groups, property managers, and retirement communities.

Unlike other robotics startups, Unix AI has not pursued the increasingly popular end-to-end VLA (vision-language-action) approach, citing limited training data. Instead, it breaks down tasks into key points and trajectories, then applies imitation learning. With only a small number of demonstrations, its robots can master new movements. Once deployed in real-world settings, they continue to refine themselves through a self-reinforcing data flywheel.

Yang, born in 2000, studied computer science at the University of Michigan and later began a PhD in the same field at Yale University before pausing his studies in 2024 to establish Unix AI. “Over the past two decades, hardware has been dominated by Chinese companies,” Yang said. “With China’s supply chain and market advantages, I saw an opening.”

During his conversation with 36Kr, Yang also shared updates on the company’s yet-to-be-released third generation of its humanoid robot, Wanda.

The following transcript has been edited and consolidated for brevity and clarity.

36Kr: Unix AI won two gold medals and one silver medal at the World Humanoid Robot Games. What came next for the company?

Yang Fengyu (YF): As soon as the competition ended, our hotline was flooded. In the second week alone, more than ten hotel clients visited our facilities.

Although our work-oriented events didn’t draw much attention onsite—we didn’t even make it onto the big screen—the results generated interest among potential clients. The preparation also improved our robots’ capabilities.

For example, the reception event required robots to lift a suitcase onto a luggage cart and then push it to a designated spot. The challenge was that the pulling direction didn’t always align with the robot’s forward motion, creating multiple hardware issues. We spent more than a month iterating on hardware before the task was removed, but the process made our robots stronger.

36Kr: Why choose hotels as the starting point?

YF: Hotel cleaning is what we consider a “quasi-consumer” skill. Once robots master cleaning, tidying, and trash collection in hotels, those atomic actions can be transferred to homes, restaurants, or retirement facilities.

Hotels also generate valuable data. Unlike industrial sites, which restrict access for confidentiality reasons, hotels allow us to feed cleaning data back into our models. They also offer higher tolerance for mistakes, since robots often work behind closed doors with less human interaction risk.

36Kr: So the medals reflected accumulated experience rather than one-off preparation?

YF: Exactly. The cleaning competition simulated a hotel room where robots had to pick up scattered bottles and boxes, which our robots already handled well.

They can also take on more complex tasks, such as removing trash bags, making beds, or cleaning bathrooms.

36Kr: Unix AI’s model is to deliver robots while simultaneously collecting data. Why pursue this strategy?

YF: We’re taking reference from Tesla. First, deploy enough robots in real-world scenarios. Then use a data flywheel to build scale.

This lowers the training threshold. We don’t even need algorithm engineers. Deployment engineers can manage it.

Just as scaling laws in large language models show how data quantity drives breakthroughs, I believe embodied intelligence can achieve the same. But it depends on how you scale.

Data diversity is more important than volume. I’d rather have one billion naturally distributed data points than a curated handful. The only way to gather natural data is through real-world deployment, not staged collection.

Scale is equally critical. In multimodal AI, training requires data in the billions. In autonomous driving, even with clean data, you need at least hundreds of thousands of cars to approach Level 4 performance. Robotics will need a comparable scale. Without tens of thousands of robots in the field, you can’t build a robust model.

36Kr: You faced challenges with the “closing the door” task during the competition but solved them quickly. How?

YF: Closing a door is inherently difficult as it involves hinges, body angles, and handle manipulation. At the venue, we discovered the door was a meter wide, larger than the 75–80 centimeters we had trained for, which broke our dual-arm closing strategy.

That night, we used virtual reality (VR) equipment to capture new data and retrained the skill. The next morning, we were first up, with no chance for further adjustments. Fortunately, we succeeded. Our imitation learning platform, UniFlex, was crucial. It can learn a new task from just five to ten demonstrations.

36Kr: Can you explain UniFlex?

YF: It’s a perception-operation decoupled model based on key-point imitation learning. We break actions into key points and motion trajectories and learn them in topological space.

It’s related to approaches like DMP (dynamic movement primitives) and VMP (variational movement primitives), which are less discussed today but are seeing a resurgence with large models.

With just a few demonstrations, a robot can generalize to similar tasks, such as closing different doors even when the handle’s design or position changes slightly.

36Kr: Many companies are betting on the VLA approach. Why not Unix AI?

YF: In the long run, I believe in the VLA approach. But for now, with limited robot data, end-to-end VLA isn’t practical.

36Kr: Some companies are combining the VLA model with tactile input, creating VTLA. What’s your view?

YF: Tactile sensing is crucial. Our UniTouch system fuses vision and tactile data to improve material recognition and contact feedback, enabling more humanlike handling.

Unlike VTLA, which compresses multiple modalities into a latent vector, we integrate tactile signals directly into our key-point framework. This enables robots to “imagine” how an object should feel based solely on visual input.

36Kr: Do your robots currently have tactile sensors?

YF: Not yet. The challenge is balancing density, durability, and cost. A single high-precision tactile sensor for one finger costs RMB 6,000–8,000 (USD 840–1,120), wears out quickly, and makes grippers bulkier. The cost-benefit ratio isn’t there yet.

36Kr: How important is hardware to Unix AI?

YF: This year marks the beginning of mass production in robotics. Hardware stability is paramount.

36Kr: Why insist on full-stack hardware development?

YF: Three reasons. First, suppliers are too slow, while in-house development gives us control. Supplier solutions also create black boxes, complicating debugging.

Second, cutting out middlemen reduces costs. Instead of buying a harmonic reducer for RMB 1,000–2,000 (USD 140–280), we use our own design, which is cheaper even for an entire joint. That’s why we can price our second-generation Wanda robot at RMB 88,000 (USD 12,300) and still maintain healthy margins.

Third, data consistency. If hardware isn’t made in-house, data from different robot generations may be incompatible, rendering models useless.

The hardest part is supply chain stability and quality control, which we’ve been focused on all year.

36Kr: Your third-generation Wanda robot appeared at the competition. How does it differ from the second generation?

YF: The third generation is designed purely for work. It’s less humanlike in appearance but more powerful, with omnidirectional wheels, higher load capacity, and better height control.

But the unit we showcased was still experimental. It was fresh from assembly when we sent it to the World Robot Conference and then the games. Its algorithms weren’t fully tested, but the results were promising.

The second-generation Wanda is stable, already in mass production, and served as our backup.

36Kr: As a CEO born in 2000, what’s your advantage?

YF: Being born after 2000 doesn’t mean much to me. What matters is that young entrepreneurs in embodied intelligence aren’t bound by old ways of thinking.

I come from a technical background. I write code myself, so I know what works and what doesn’t. This is a technology-driven field, and young people are the driving force.

36Kr: How did you start this journey so young?

YF: I’ve always wanted to start a company. I also studied finance and business. My PhD advisor is also an entrepreneur, and he encouraged me.

In academia, you often create problems to solve problems. I prefer tackling real-world issues. Over the past two decades, hardware has been dominated by Chinese companies. With China’s supply chain and market advantages, I saw an opening.

36Kr: Your team is young. How do you attract experienced veterans like your chief scientist, Wang Hesheng?

YF: It’s about complementarity. A group of veterans working together will still follow veteran playbooks. We bring energy and sensitivity to the frontier of technology. The veterans bring industry experience and resources. Together, the combination creates powerful chemistry.

KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Fu Chong for 36Kr.