Fast and pragmatic. That’s how Liu Fang, founder of Amigos Robots and former head of Xiaomi’s smart driving division, is often described. He started Amigos in September 2024, and a year later the first batch of robots was already running on customer production lines.
According to information obtained by 36Kr, Amigos raised both seed and angel funding this year. The seed round was co-led by Anker Innovations and Xinglian Capital, with Sunwoda and K2VC participating. The angel round included CICC Capital, Peakvest, and Xinglian Capital. In total, the company has raised nearly RMB 200 million (USD 28 million), with Yuefeng Capital serving as financial advisor.
“In China’s market, there are no real technical secrets. What matters in the end is continuously solving customer needs, and that’s how you build customer loyalty,” Liu told 36Kr.
Liu joined Xiaomi in 2012 during its startup phase. Over 13 years, he oversaw smartphone systems, artificial intelligence hardware, and autonomous driving. Those experiences, he said, shaped his cost discipline, customer focus, and fixation on efficiency. That mindset now defines his startup, especially in choosing the right use cases for robots.
Before founding Amigos, Liu conducted extensive market research, breaking the problem into three criteria: clear demand, significant improvement through AI, and measurable return on investment (ROI).
He believes embodied intelligence in industrial settings is not meant to replace automation, but to fill gaps where labor costs are high or automation falls short.
Although labor costs in China are relatively low, factories face worker shortages and high turnover. As production shifts to small, fast-iterating batches, traditional automated lines become costly to reconfigure. Embodied intelligence, which can learn faster, offers a more flexible alternative.
Amigos targets complex manufacturing steps such as sorting, assembly, and inspection, where traditional automation is inefficient and manual labor is expensive.
The economics are straightforward. A factory worker in coastal China earns about RMB 6,000–7,000 (USD 840–980) per month, or roughly RMB 80,000–100,000 (USD 11,200–14,000) a year. For a three-shift operation, one workstation can cost RMB 200,000–300,000 (USD 28,000–42,000) annually.
Following that math, Liu priced each Amigos robot at around RMB 200,000. He found that if the payback period stays within 12–18 months, factory owners view embodied intelligence as worth the investment. Beyond that window, hesitation grows.
As an entrepreneur born in the 1980s, Liu has a nuanced view of competition between tech giants and startups.
Some argue AI’s endgame will be dominated by large technology companies. Liu sees manufacturing’s modest margins as a deterrent, giving startups room to operate.
In the long term, he also sees opportunities abroad. As geopolitics pushes Chinese manufacturing capacity overseas, domestic embodied intelligence companies could ride the wave of industrial migration.
36Kr sat down with Liu to discuss how Amigos identifies real-world scenarios for embodied intelligence and his outlook on the sector.
The following transcript has been edited and consolidated for brevity and clarity.
36Kr: Why choose factories as your starting point for embodied intelligence?
Liu Fang (LF): Before starting the company, I spent a lot of time on market research. The key was to find a scenario with clear demand and an ROI that makes sense.
Honestly, everyone wants to take the B2C direction. It sounds exciting and has a bigger imagination space. But our research showed that, at present, the technology and cost just aren’t there.
Take household robotics, for example. In China, domestic helpers are still affordable. And beyond technology, there’s a human factor: emotional and ethical barriers in interacting with machines.
For example, when I traveled to Japan, I noticed that even though Japan’s service industry is advanced, restaurants that value efficiency use self-service ordering machines. But customers willing to pay more for premium service still prefer human interaction.
In contrast, the logic for industrial applications is clear. A workstation costs about RMB 100,000 (USD 14,000) per worker per year, or RMB 300,000 (USD 42,000) for a three-shift operation. If one robot, priced at RMB 200,000 (USD 28,000), can replace two to two-and-a-half workers, the client breaks even within one to two years. That’s why factories can make quick purchasing decisions.
36Kr: Among various industrial applications, why target 3C (computer, communication, and consumer electronics) manufacturing?
LF: We screen scenarios using three criteria: clear demand, measurable ROI, and AI’s ability to deliver significant improvement.
3C factories are highly labor-intensive, with tens of thousands of workers and densely packed workstations, which makes deployment easier. Labor accounts for 12–15% of total costs, giving these factories both the incentive and the budget to upgrade.
36Kr: You’ve predicted that embodied intelligence will handle routine factory tasks by late 2026, while complex operations may take another year or two. How did you estimate that?
LF: The timeline is based on data accumulation and adaptation cycles.
Once we reach tens of thousands of hours of recorded data, robots will be capable of handling flexible assembly tasks. Our goal this year is to collect several thousand hours of data and to reach five figures by the end of next year, establishing the foundation to train robots in complex physical tasks.
36Kr: How do you evaluate technical barriers in embodied intelligence?
LF: In China, no one holds a monopoly on technology. The real edge is how quickly you solve real customer problems and build lasting relationships. Our competitive moat is speed, making the right decision fast.
Rather than creating all-purpose robots, we focus on making robots that can learn fast. The goal is for a robot to master a new workstation quickly.
We adopted a strategy that utilizes first-person video data early on, whereby workers wear cameras, allowing robots to learn by watching real operations. This minimally disrupts workflow while collecting highly relevant training data.
At Amigos, the ratio of video to machine-generated data is about six to one. Most training data come from videos, with a smaller portion from real-world fine-tuning.
It’s like an apprentice learning from video tutorials and then doing a few practice rounds before taking over. Our goal is to cut deployment time for new workstations from months to under a week.
36Kr: So it’s not about making robots that can do everything from the start?
LF: Exactly. We believe in the idea of a data flywheel, but we’re not chasing scale for its own sake. We care about making the flywheel spin effectively.
Embodied intelligence needs specialization and speed in solving focused problems. Once a robot masters several tasks, it gains generalized capabilities within that environment.
36Kr: How is Amigos training its core models? What’s your take on reinforcement learning?
LF: We use a vision-language-action (VLA) framework for higher-level decision-making and planning, while reinforcement learning handles fine control.
Crucially, our reinforcement learning happens on real machines, not simulations. Simulations can’t fully capture tactile feedback or environmental nuances.
On-machine reinforcement learning helps us in two ways: improving fine-grained tasks like millimeter-level grasping and enhancing self-correction when errors occur.
For example, simulations can show how a motion should be executed, but only real-world practice can tell whether it’s done right. Training in real environments makes robots more reliable.
36Kr: Some say embodied intelligence risks data security and fault tolerance in industrial use cases. How do you handle that?
LF: Our current clients, which are typically midsize factories with 10,000–20,000 workers, are willing to share data, providing enough scale for cross-factory model training.
Once our base models are robust, we can work with more data-restricted clients by fine-tuning locally.
Also, low fault tolerance isn’t unique to factories. Even bubble tea shops run at high speeds. But factory environments are structured and consistent, which actually makes them more suitable for embodied intelligence deployment.
36Kr: How about cutting-edge technologies like tactile sensing and world models?
LF: Embodied intelligence is itself deep technology. Everyone’s still searching for stable, scalable engineering paths.
We track advances like VLA models, world models, and multimodal sensing. We do that not just for their novelty but for their problem-solving value. For example, we’re exploring how multimodal sensing can improve precision in assembly.
36Kr: How is commercialization progressing?
LF: We’ve already deployed robots with three key accounts. They have been running alongside human workers for some time, and clients are now considering larger purchases. Progress is faster than we expected.
36Kr: Why start a company now, and why focus on embodied intelligence?
LF: Robotics, like automotives, is a traditional field with decades of groundwork. AI gives it exponential growth potential, just as it did for autonomous driving. To me, embodied intelligence is a clear, inevitable opportunity.
At large corporations, you can see the tech trends clearly, but decision-making tends to be cautious. For people who want to drive change, that means missed opportunities. So I chose to build something myself.
36Kr: How did the company name come about?
LF: It comes from the Spanish word for “friends.” Our vision is to make robots companions to humans, and “Amigos” sounds friendlier in Chinese.
36Kr: Many embodied intelligence founders are in their 20s or early 30s. As a veteran, how do you see that?
LF: Yes, I’m older, but that’s not a disadvantage. Having lived through multiple tech cycles, I’ve gained perspective and humility.
In every tech wave, the founders who endure are those with prior experience. In industries that require complex supply chain coordination, experience matters even more.
We also hire young engineers who are immersed in the latest technologies, so our team balances experience and innovation.
36Kr: Does Amigos plan to expand overseas?
LF: Yes. There are two drivers, clients relocating production abroad, and higher ROI in overseas markets.
For instance, a Hungarian client told us it would consider embodied robots priced under EUR 150,000 (USD 172,975).
China, as the world’s largest manufacturing base, offers the richest demand and scenarios. Our plan is to perfect our technology, products, and service at home, then follow Chinese clients abroad before expanding to local customers.
36Kr: What’s your take on embodied intelligence being a potential bubble? How should we gauge the market next year?
LF: Embodied intelligence is a fast-moving, capital-fueled sector, so there’s naturally some hype. The best indicator is how top players perform.
But the sector will mature, and that’s inevitable. Value will crystallize over time. Rather than chasing valuations, we prefer to build on healthy, sustainable cash flow.
Next year’s trajectory will depend on one thing, real commercialization.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Fu Chong for 36Kr.