Five Embodied Intelligence Founders on Business Logic, Breaking Through Bottlenecks, and Differentiation Strategies | Yunqi Capital Practitioners --- *Note: This is a title/header translation. The full article content was not provided for translation. If you share the complete Chinese text, I can translate the entire piece following all specified guidelines.*

AI, embodied intelligence, going global... in 2024's tech and venture capital circles, few people could avoid crossing paths with these terms. As the year draws to a close, the "Yunqi Capital Pragmatists" column is launching a year-end special series. We sat down with Yunqi Capital portfolio founders working deep in these trenches to discuss the changes and opportunities in these sectors and directions this year, hoping to bring you on-the-ground perspectives and insights from those building in the trenches.

In the second installment of "Yunqi Capital Pragmatists: Year-End Season," we turn our attention to embodied intelligence. (For the first piece in this series, click here 👉🏻 Yunqi Capital in Conversation with MiniMax's Junjie Yan: Looking Back at 2024, What's Changed and What Hasn't in AI Entrepreneurship | Yunqi Capital Pragmatists)*

In 2024, now just behind us, AI application and deployment drew significant attention. As one of the more imaginative directions for AI deployment, embodied intelligence concepts and the startups built around them became hot topics in tech and VC circles. Meanwhile, relatively more established categories like industrial robots and service robots found renewed possibilities under the boost of new technologies.

But on the long march toward AGI, how can technological innovation proceed in parallel with commercial path exploration? How large is the gap between current embodied intelligence technology and the generality and generalization capabilities people are hoping for? In late 2024, we spoke with five founders in embodied intelligence and related fields, enriching our understanding through their varied practical experiences. This issue excerpts some of the highlights to share with you.

• Founders participating in the conversation (in alphabetical order by surname)

Xuan Liu, Technical Partner, Yunqi Capital

A provider of high-level intelligent driving solutions, one of the few domestic companies to achieve mass production of "end-to-end" driving systems, with 30,000+ production vehicles in 2024

Cewu Lu, Co-founder, Noematrix

An embodied intelligence startup that completed three funding rounds in 2024, dedicated to foundational models and systems for embodied intelligence

Lianyang Ma, Head of Perception Technology, Astribot

An embodied intelligence startup that released its first-generation AI robot, the Astribot S1, in 2024 and began market pre-sales

Bin Wan, COO, Keenon Robotics

The world's largest service robot supplier by shipment volume, with #1 market share in food service robots

Suibing Zheng, Founder & CEO, RealMan Robotics

The industry's first provider of ultra-lightweight humanoid robotic arms, with annual production capacity of 20,000+ units

01 From Frontier Technology to Commercial Deployment: How to Get There?

Cewu Lu, Noematrix

Our commercialization logic rests on a fundamental consensus: the endgame for robots is the ability to perform all kinds of manipulation tasks, representing a massive market and demand. But in getting to that endgame, we need to find a path that can simultaneously advance technical generalization and achieve scale and commercial viability.

Along the way, we identified something key: real-world tasks are endlessly varied, and it's impractical to design independent workflows for each one. But all tasks can be broken down into individual skills. After analyzing over 1,000 commercial cases, we found that robots actually only need about 100 skills — inserting, pulling, twisting, unplugging, pushing, and so on — and these skills are generalizable across many scenarios. So you don't need to throw too many engineers at each scenario application. Grasping, for example, is essentially the same whether it's in a hospital, a kitchen, or a factory.

So what we're doing is leveraging our AI technology strengths to build a system, essentially providing a Photoshop-like toolkit that enables different skills to be arranged and combined. On this path, we began commercialization in August 2024, and order demand grew rapidly in a short period. Because this is a path that can drive both commercial viability and generalization in tandem.

Xuan Liu, Yunqi Capital

Yunqi Capital's commercialization path is to partner with automakers, integrating our intelligent driving system into their vehicles for the consumer market. The traditional rule-based, modular Robotaxi commercialization model has problems — it's too expensive, and the amount of data you can collect is limited. So we quickly began laying out pre-installed mass production solutions, partnering with automakers to launch vehicles equipped with Yunqi Capital's high-level intelligent driving system. These vehicles drive on roads every day, generating large volumes of high-quality desensitized data that help our algorithms iterate rapidly.

Yunqi Capital has achieved what it has today because we made forward-looking technology investments two to three years ago. For example, we began developing solutions that don't rely on high-definition maps in 2020, and conducted the first road tests of our end-to-end model in 2023. Our early investments in mapless and end-to-end technology were later validated as becoming industry mainstream. Currently, there are only three suppliers in the industry capable of deploying high-level intelligent driving within cities, and Yunqi Capital is one of them.

We're also continuously strengthening our engineering capabilities, having established partnerships with multiple automakers. In 2025, we expect to have more than ten vehicle models hit the market.

Bin Wan, Keenon Robotics

Keenon Robotics' journey of more than a decade has been a continuous process of innovation and exploration. The first sector we deployed in wasn't actually the service industry we focus on now — we explored for several years before pivoting into it and beginning to deliver efficiency gains for the service sector. Then in 2016, we launched the world's first service robot and moved into mass production. In 2018, we entered into large-scale strategic cooperation with Haidilao, and subsequently expanded into more service scenarios including hotels, elder care, and education.

Against the backdrop of rapid AI technology development, we're constantly integrating the latest AI technologies into robotic applications. Our advantage lies in having tens of thousands of machines running out in the world every day, solving concrete problems in many commercial scenarios. Now we can leverage the latest AI technologies to make tasks more closed-loop. In hotel delivery scenarios, for example, robots need to improve their recognition and judgment capabilities regarding elevators, rooms, and items being picked up or delivered. To this end, we're combining large models, embodied intelligence, and other technologies to continuously iterate and enhance these capabilities. Keenon Robotics' embodied intelligent service robots already possess stronger generalization capabilities, able to flexibly adapt to different scenario requirements.

So our core is applying the latest technologies to strengthen our own commercial closed loop.

Suibing Zheng, RealMan Robotics

Moving from frontier technology to commercial deployment requires effort on multiple dimensions. Between 2017 and 2018, my team and I closely tracked market dynamics and deeply researched actual application scenarios in industries like unmanned retail and food service, from which we identified strong market demand for lighter, more general-purpose robotic products. This led us to settle on the R&D direction of ultra-lightweight humanoid robotic arms.

Our commercialization process also closely revolves around the three main bottlenecks of embodied intelligence: manipulation capability, generalization, and cost. For example, our ultra-lightweight humanoid robotic arm is designed with full reference to the length, thickness, flexibility, and load-bearing capacity of the human arm. From appearance to performance, we strive to achieve "if a human can do it, it can do it," while leveraging the "ultra-lightweight" advantage to make it adaptable to diverse scenario needs.

In terms of improving robotic generalization capabilities, we also established a dedicated AI Lab in recent years. We break down generalization into "cerebellum" and "cerebrum" functions, achieving major breakthroughs in object recognition, localization, and manipulation. Our robots can skillfully operate delicate items without additional training, better adapting to diverse living scenarios.

Cost control is also critical to commercialization. By optimizing production processes and meticulously refining every link from supply chain management to manufacturing processes, we've successfully kept the GEN72 robotic arm's price under 10,000 RMB without sacrificing product performance, a move that has greatly enhanced the product's market competitiveness.

02 The Generalization Bottleneck in Embodied Intelligence: How to Solve It?

Lianyang Ma, Astribot

In recent years, AI has evolved significantly across text, image, and voice dimensions, but there are still obvious deficiencies when it comes to interacting with the physical world. For us, what the embodied intelligence track lacks is real-machine data, or rather, training models based on real-machine data.

Embodied manipulation data is essentially human-like or derived from humans. How can robots highly faithfully reproduce human movements? Solving this problem makes the collected data more conducive for models to learn from humans. On this point, a goal we set when designing our product was that hardware performance should match or even exceed human capabilities, something we've already achieved. So at this level, we believe that in embodied intelligence humanoid manipulation, hardware determines the ceiling.

Of course, how software or algorithms gradually approach that ceiling is equally important for us. We're currently iterating network models in parallel, validating the optimal synergy between algorithms, data, and the physical body — it's just that relative priorities differ at different stages.

Generalization is the ultimate goal that embodied intelligence must achieve, encompassing both generalization of manipulation skills and generalization of manipulation objects. Currently, embodied intelligence has basically achieved generalization with limited objects and limited skills — for example, a robot making coffee can generalize across different types of coffee capsules and cups, and if you need to select capsules of different colors, we can also achieve successful grasping and related operations. The next step is having limited manipulation skills but broadening and generalizing the range of manipulation objects. Continuing to evolve requires achieving generalization across unlimited skills and unlimited object dimensions — this is the omnipotent embodied intelligence everyone wants. This process requires joint efforts across data, models, computing platforms, and government industry policies. Technology is currently developing rapidly, with tremendous room for imagination.

Cewu Lu, Noematrix

A good company needs to establish a flywheel of algorithms, data, and deployment scenarios — this is crucial. Over the past year, we've focused on building such a flywheel. So you can see we have some unique data collection solutions. With data, you can automatically train a model. In the model validation process, it gets tested against your scenarios. Only if your model is better can you try more scenarios; otherwise, if the model performance isn't there, you're stuck testing in a very limited set of scenarios and definitely can't expand.

Suibing Zheng, RealMan Robotics

I believe the biggest bottleneck for embodied intelligence deployment is application. The first landing moment for end-to-end models was Tesla's FSD, and that moment came based on large volumes of real-machine road test data generated by users driving vehicles. I think we can learn from autonomous driving's experience, accumulating data and optimizing through iteration in real applications, because only large-scale practical application can generate sufficient data to drive continuous progress in embodied intelligence.

03 The Track Is Crowded — How to Differentiate?

Xuan Liu, Yunqi Capital

Currently, many competitors still lean more toward the traditional rule-based, modular Robotaxi model. This model, particularly in the US, has been gradually abandoned by the industry. Recently, General Motors also announced it would no longer invest in Cruise's autonomous taxi project, believing that Tesla's "end-to-end" direction is the correct one.

This also means autonomous driving has entered a new era of data-driven development. On this technical path, we are the first Chinese company to truly achieve mass production and deployment of "end-to-end" intelligent driving, and our first model sold over 30,000 units within four months of launch, with data growing very rapidly. With large volumes of data secured, in 2025 we will also deploy the next generation of end-to-end technology, the Vision-Language-Action (VLA) model, in vehicles. These are relatively leading developments in the industry.

Going forward, we hope to migrate data from the intelligent driving vehicle domain to the broader physical world, building general artificial intelligence for the physical world. Overall, in the artificial intelligence industry, we are more advanced on the "intelligence" dimension.

Bin Wan, Keenon Robotics

In the long run, the service robot sector remains a vast ocean of stars; currently, penetration is still extremely low, and applications are still very few.

Internally, one important dimension we're working on now is penetration. Scenario by scenario, even customer type by customer type. The food service industry, for example, sounds like just one sector, but it can actually be divided into many small sub-sectors like hot pot, barbecue, Japanese cuisine, Korean cuisine, and so on — with significant differences between them. Similar situations exist in healthcare, where different departments have notably different patient care needs. Including going global: different national markets vary greatly. For instance, elevator control for hotel robots, common domestically, is difficult to install in certain overseas countries, meaning the same product needs to be re-adapted and calibrated for different scenarios and markets.

Another dimension is expansion. The service industry should theoretically be very broad, encompassing what you might call 360 trades, but not many sectors can be scaled. Beyond hotels, food service, and cleaning, we haven't yet seen more sectors with comparable volume. So we need to continuously find new breakthrough points and expand into new verticals.

Only through the combined effect of penetration and expansion can we continuously advance toward that ocean of stars. As technology continues to evolve, Keenon Robotics' embodied intelligent service robots not only possess stronger decision-making capabilities and adaptability, but can also achieve closed-loop operations in complex environments. For example, Keenon Robotics' dual-arm W3 hotel robot can independently take elevators to address overseas elevator control challenges, and place delivered items on tables, providing users with a more complete delivery experience. Going forward, we will drive intelligent transformation across more industries, expanding the boundaries of commercial application.

So persisting with long-termism, staying grounded — this is the more sustainable choice for development.