Yunqi Capital Attent!on Podcast | Founders Up and Down the Supply Chain: From Embodiment to Intelligence, How Is AI "Awakening" Robots?

Before the "robot dream" of machines that can learn, think, and move freely becomes reality, what niche entry points and scenarios can unlock the technical advantages of embodied intelligence? What chasms must be crossed to truly achieve the generality and generalizability of embodied intelligence? From an early-stage investment perspective, what are the business models and strategic logic for embodied intelligence?

At a recent Yunqi Cloud Summit DemDay event focused on embodied intelligence, three Yunqi portfolio founders operating at different points of the embodied intelligence ecosystem — RealMan Robotics, Astribot, and C12.ai — joined Yunqi Capital partner Chen Yu for a lively discussion on these very questions.

This edition of Yunqi Attent!on: Embodied Intelligence Series brings you the highlights. Both transcript and full audio available — read on or tune in.

Scan the QR code above or follow "Attent!on" on Xiaoyuzhou to listen to this episode

· Guest introductions:

Chen Yu, Partner at Yunqi Capital

Deeply invested in AI and robotics, with representative investments including MiniMax (large model unicorn), Keenon Robotics (leading restaurant robotics company), DeepRoute.ai (top autonomous driving company), PingCAP (open-source database unicorn), and others.

Zheng Suibing, CEO of RealMan Robotics

RealMan Robotics is a global leader in ultra-lightweight humanoid robotic arms, with over a decade of experience in collaborative robotic arms and core components R&D. The company serves more than 4,000 enterprise clients and has accumulated thousands of benchmark cases.

Lai Jie, CEO of Astribot

With 16 years of robotics R&D experience, Lai was the first employee at Tencent's Robotics Lab and led Baidu's "Xiaodu Robot" team. Astribot is dedicated to building general-purpose manipulation robots, with its first AI robot's complete body scheduled for release in August.

Chen Zhigang, CEO of C12.ai

Previously served as Chief Digital Officer at WuXi AppTec, head of Tencent's medical big data lab, and chief architect at Alibaba Health. C12.ai is building embodied intelligence for laboratories, providing next-generation intelligent solutions for pharmaceutical R&D labs.

Below are excerpts from this episode

01 Humanoid Robots Steal the Show at WAIC: What It Tells Us

Chen Yu, Partner at Yunqi Capital

The biggest eye-catcher at this year's WAIC (World Artificial Intelligence Conference) was definitely the "Eighteen Arhats" at the entrance. You could see so many humanoid robots on display — a real blossoming of the field. Why this year? I think a major reason is that technology has reached a point where computing power, large models, machine learning theory, and hardware have all prepared the ground for embodied intelligence. So suddenly you're seeing so many companies launching humanoid robots all at once.

Zheng Suibing, CEO of RealMan Robotics

I think embodied intelligence is booming because it fulfills people's imagination of the future. We want to build something with capabilities equal to our own, or even something god-like. From that perspective, AI landing on robots is an inevitable destination.

Over the years, efforts by companies like Boston Dynamics and Tesla gave robots manipulation capabilities before the AI explosion. Then large models emerged and solved a critical capability for robots: nested logical reasoning. Tell a robot to open the fridge and get a Coke, and it knows the logic — walk over, open the fridge door. Ask it how to open the fridge door, and it'll tell you: first see the handle, grab the handle, pull it open. You see it has nested logical reasoning. And nested logical reasoning is precisely what solves the generalization problem that has plagued robot deployment in real-world applications.

There are three bottlenecks for robots entering daily life: generalization capability, manipulation capability, and cost. Two of these have now been solved.

02 Getting Humanoid Robots to Move Freely: Two Gaps to Cross

Lai Jie, CEO of Astribot

The embodied intelligence industry actually faces two gaps: one between edge AI and the robot body itself, and another between embodied intelligence and large models. To achieve a really smooth demonstration, first the body needs to reach a certain level of performance and stability, then edge AI needs to give it generalization and adaptability — only this complete combination allows for natural demonstration.

Of course, for us, what we're doing now still leans toward data-driven approaches. Having robots complete current tasks under adaptive and generalizable capabilities — I see this as a two-way street. Simply adding some AI capabilities on top of existing robots exposes many flaws in demos.

So we're now focused on solving the gap between the body and edge AI first, designing AI models and the body together through a unified top-level architecture. At the same time, we're also working on AI agents and even pushing forward world models, building toward the robot form that people imagine appearing by their side.

03 Embodied Intelligence + Drug Discovery: How Does It Land?

Chen Zhigang, CEO of C12.ai

In pharmaceutical R&D scenarios, many tasks are difficult to complete with an end-to-end model. For example, thousands of drug targets have been discovered globally, yet we're actually searching through a compound space of 10^60 possibilities, requiring massive experimentation during synthesis. Much of this experimentation has never been done before, and the exploratory process involves a hierarchical structure. This hierarchical structure is the specialized brain we're now designing.

These brains essentially decompose very complex tasks, using an approach called long-range. The benefit of this approach is that by decomposing complex problems, they become controllably solvable. Second, interpretability is enhanced — unlike end-to-end models where problems are hard to trace, in long-range mode we can make timely adjustments and interventions to task design.

Meanwhile, the落地 of embodied intelligence also depends on the development of the entire ecosystem. In the future, collaboration between the software layer and hardware control layer can help us better explore落地 space.

04 Is the Humanoid Robot the Ultimate Form of Embodied Intelligence?

Lai Jie, CEO of Astribot

I actually worked on bipedal robots until 2021, but thought carefully about this question before starting my company. First, there's the problem of usable time. We believe the first element of commercial value in robotics is sufficient working time, and clearly this is currently a bottleneck for bipedal designs. So we believe manipulation capability has greater value, while a mobile base brings stability and longer usable time. That's why we chose to develop comprehensive manipulation capability first.

Speaking of endgame, I think if energy technology advances, humanoid forms could develop significantly. Or in certain special scenarios — mountainous terrain, for instance — there's genuine demand for humanoid forms.

Chen Yu, Partner at Yunqi Capital

Everyone's dream is certainly to build humanoid forms, and the reasoning is sound — after all, our environment is built around humans. But from technical and cost considerations, whether to use bipedal design isn't mandatory. From a practical standpoint, manipulation capability matters more. Though the ultimate form will still lean humanoid.

05 What's the Business Model for Embodied Intelligence? What's Our Investment Logic?

Chen Yu, Partner at Yunqi Capital

Embodied intelligence should ultimately replace or assist humans in completing work, so the business model benchmarks against human labor. In certain application scenarios — like pharmaceutical labs where C12.ai is focused, or other high labor-cost environments — if robots can be cheaper than the corresponding human labor, the market will pay.

From a more endgame perspective, everyone eventually wants a capable personal assistant, but that business model still needs considerable time. I think this is somewhat similar to the development pattern of autonomous driving — around 2014-2016, we could already see autonomous driving prototypes running on roads. I believe embodied intelligence is on the same path now: we build the prototype first, then spend 5-10 years driving down costs across every link in the chain.

From an investment perspective, embodied intelligence is an extremely complex industry spanning software, hardware, and multiple layers — components, bodies, data collection, simulators, and so on. As investors, we currently want to be as comprehensive as possible in our coverage.

Additionally, technical approaches in embodied intelligence haven't converged yet — there are end-to-end machine learning approaches, large model plus reinforcement learning approaches, and others. In these early technical stages, we're also making multiple bets on companies with different technical approaches. In the future, approaches may not necessarily converge — each technical path may have its own value, capable of producing embodied intelligence with high cost-performance and high functionality.

Podcast Timeline

01:29 Topic and guest introductions

06:24 Humanoid robots steal the show at WAIC: large models unlock robot generalization capability

12:27 Three bottlenecks for robots entering daily life: generalization capability, manipulation capability, cost

13:46 Which industries are actively embracing embodied intelligence?

14:56 How to get humanoid robots moving? Two gaps to solve: edge AI vs. robot body; embodied intelligence vs. large model

16:14 What's the commercial value of embodied intelligence? AI robot entrepreneurs weigh in

17:43 Is the humanoid robot the ultimate form of embodied intelligence?

20:12 Is the humanoid robot the ultimate form of embodied intelligence?

26:03 Industry ecosystem outlook: upstream-downstream collaboration or large platform "do-it-all"?

27:14 Business models and investment logic from an early-stage perspective