AgiBot's Chief Scientist: "The Biggest Misunderstanding of Embodied AI Is Treating It Like a Large Language Model" | BlueRun Ventures Family Headline

This article is republished with permission from Waves (ID: waves36kr), by Lili Yu

The embodied intelligence sector is currently caught in something of a spectacle. On one side, investors like Allen Zhu are making high-profile exits from what they see as a bubble. On the other, a direction that many investors had already written off by late last year as having "closed its betting window" keeps getting reactivated by blockbuster funding rounds. Among these are not only some newly entrant companies with autonomous driving backgrounds, but also several firms whose valuations had already skyrocketed last year. AgiBot, Tencent's latest bet, is among the most closely watched. And this marks Tencent's first investment in the embodied intelligence space. BlueRun Ventures was an early investor in AgiBot and has now added to its position for three consecutive rounds. As early as 2023, the company — founded by Taihua Deng, former president of Huawei's computing product line, and "Huawei Genius Youth" Zhihuijun (Peng Zhihui), among others — set a record by closing a 300 million RMB angel round just one month after incorporation. Not stopping there, it reached a $1 billion valuation within months, becoming the world's fastest embodied intelligence company to achieve unicorn status. Among China's top-tier embodied intelligence companies, AgiBot is unquestionably the most audacious and high-profile. The market has never been short of noise about it. Following the March release of its first general-purpose embodied foundation model, AgiBot announced a partnership with embodied intelligence company Physical Intelligence (Pi). The matchmaker was none other than Jianlan Luo, who had just been announced as AgiBot's chief scientist. Luo previously conducted research at Google X and Google DeepMind. During his postdoc at the Berkeley Artificial Intelligence Research Lab (BAIR), he was a core member of the team led by Professor Sergey Levine — a leading figure in deep reinforcement learning and one of Pi's co-founders. On why he joined AgiBot, and on the many misconceptions about the embodied intelligence sector, Luo shared his thoughts with several media outlets. Below are excerpts from those conversations, edited and compiled by Waves.

On why he joined AgiBot, and on the many misconceptions about the embodied intelligence sector, Luo shared his thoughts with several media outlets. Below are excerpts from those conversations, edited and compiled by Waves.

Question 1: Because of Allen Zhu's exit, many people now believe embodied intelligence is already in a massive bubble.

Jianlan Luo: A bubble, by its nature, also means attention and resources — it's a bet placed in advance. A flood of resources rushes in, and at some point expectations fall short, so things start coming down, and maybe rise again later. This is completely normal.

Every technology paradigm shift goes through this. Autonomous driving did too. Since Waymo started in 2016, it has only now truly reached the dawn of commercial viability and real-world deployment. Embodied intelligence is an even more complex, systems-level endeavor, which means it requires a longer period of technical accumulation rather than breakthroughs through sheer compute or model scaling.

Question 2: Is the large language model the most critical variable driving embodied intelligence's rise?

Jianlan Luo: The biggest misconception outsiders have about this industry is over-analogizing embodied intelligence to the large model paradigm.

There are similarities, and some large model techniques can transfer to embodied systems and robotics. But they cannot be simply equated.

For instance, an LLM being 50-60% accurate can still be useful. Because you have a human brain — if ChatGPT tells you to drink pesticide, you won't, because you can judge for yourself. But on a robot, that accuracy level is completely useless.

Imagine your household robot smashing a cup on your coffee table every three hours, throwing your phone at a window, or a coffee-delivery robot randomly spilling coffee every 20 minutes. Like autonomous driving — compared to ten years ago, it's a completely different species now, with very high success rates, yet people still demand more. Because every such failure has physical-world consequences.

So using the large model timeline to analogize embodied intelligence underestimates the unique challenges of operational intelligence and action intelligence.

Question 3: As things evolve through different stages, how will the importance shift between the software side represented by large models and the hardware side involving manufacturing?

Jianlan Luo: Software and hardware are equally important. Right now, software hasn't converged to a single point, hardware hasn't converged either, and the industry has no consensus on how to integrate the two.

Question 4: What are the most critical bottlenecks in software and hardware respectively right now?

Jianlan Luo: On the software side, while large models are powerful, they still lack long-term memory. Cross-task attempts, hierarchical control, and real-time feedback all remain difficult problems. Whether to use simulation, how much real data versus synthetic data is needed — none of this is settled. Including whether to use RL, because applying RL in the real world brings challenges around sample efficiency, training stability, and generalization.

On the hardware side, some high-performance platforms are still very costly. Some sensor feedback isn't fine enough — tactile sensors, for example, haven't reached maturity. Reliability also has significant room for improvement.

There are also many robot bodies and solutions, including actuators. I don't think one body will solve all problems going forward. Rather, there will be several relatively standardized bodies tailored to different industries, with corresponding solutions.

Question 5: The data problem seems to be the most controversial, but it also looks like a chicken-and-egg dilemma.

Jianlan Luo: Right, it looks like a loop where head and tail chase each other. Without sufficient data, it's hard to deploy robots to the real world. But without deployment, you can't get that data.

But imagine if 1,000 robots were working at Starbucks, making and delivering coffee 24/7 — the data flowing back in a month would exceed the scale of any robot dataset we've seen.

And robots differ from cars in another way. With cars, if you don't have 100% confidence, you can hardly put them in the real world, because safety requirements are so strict.

But robots can start in some enclosed or semi-enclosed spaces, operating at maybe 70-80% confidence, and that allows more data to flow back to improve the system.

Question 6: Autonomous driving also had extensive discussions about data in its early days.

Jianlan Luo: When autonomous driving started in 2016, there was also much debate due to data scarcity. But now there's too much data. Tesla disclosed 50 billion miles of on-road data last year — their data centers can't even hold it all. So we shouldn't worry about whether there's enough data, but rather what algorithm designs we need to better harness it. Therefore, embodied intelligence companies that control products and ecosystems and have the capability to deploy robots themselves will have massive first-mover advantages.

Question 7: In your view, is the full-stack, hardware-software-integrated approach necessary? Some companies just want to focus on the body.

Jianlan Luo: In autonomous driving's early days, there were also companies focused purely on the "brain." But now OEMs are all building their own autonomous driving capabilities. Ten years ago, when drones were hot, a wave of drone companies emerged in both China and the US. American companies said they wouldn't do hardware — I recall Intel opened over 20 labs in the US just working on drone navigation and such. Of course, this was also because the US lacks manufacturing and supply chains, so they could only do the brain. But now you don't remember any of them, because they no longer exist. The name we remember now is: DJI.

While doing only the brain can work in combination with hardware, I believe the full-stack route with iterative hardware-software integration will ultimately prevail.

Question 8: AgiBot already has CTO Zhihuijun, and Yao Maoqing, executive director of AgiBot Research Institute, also has a technical background. Is there a reporting relationship between you, and how will you divide responsibilities?

Jianlan Luo: Internally, we're a relatively flat, highly collaborative team. Zhihuijun has deep accumulated expertise in systems engineering. Director Yao maintains strategic oversight of the big picture. I'll focus more on pushing algorithm roadmaps and integrating external technical ecosystems.

We're in a parallel, complementary relationship, with more emphasis on consensus-driven, project-oriented collaboration.

Question 9: What was the background for AgiBot's decision to partner with Pi (Physical Intelligence)?

Jianlan Luo: First, AgiBot and Pi share many aligned philosophies — both emphasize the importance of real data, both push embodied intelligence toward practical deployment from real-world foundations. That's the broad context for the partnership.

Additionally, it was founded by pioneers in embodied intelligence including Professors Sergey Levine and Chelsea Finn, and is currently among the best embodied intelligence companies internationally.

Question 10: Among embodied intelligence startups, AgiBot has also been operating with something of an ecosystem approach, quite like running a startup with big-company methods. Is this intentional?

Jianlan Luo: We believe the complexity of embodied intelligence far exceeds what any single company can bear. So we emphasize open collaboration. On one hand, we help external companies achieve their own iterations; on the other, we bring their capabilities into our ecosystem.

Question 11: Why hasn't an OpenAI-like, defining star company emerged in the embodied intelligence space yet?

Jianlan Luo: Because the industry hasn't converged on highly deterministic technical solutions, so no company is far ahead with strong discourse power.

Question 12: People see many impressive robot demo videos, but ultimately they're all human teleoperated. How do we achieve autonomous decision-making?

Jianlan Luo: The difference between autonomous decision-making and teleoperation is like thinking you're chatting with ChatGPT, but actually there's another person typing on another computer behind it — completely different things.

The essence is the robot's analysis and modeling of uncertainty, then converting that into executable action chains. For a robot, if the position shifts slightly, or the color changes, it no longer matches what it memorized. The generalization capability of this perception-prediction-generation mechanism is the most critical technology.

Question 13: Recently, embodied intelligence manufacturers have been eager to show off skills in executing long-horizon, complex tasks, with each company's skill points differing somewhat. How do you define long-horizon and complex tasks?

Jianlan Luo: "Long-horizon" is a relatively subjective term. We're more focused on whether a task involves complex sequential dependencies, and its generalization capability, rather than some absolute condition where one minute is long-horizon and under one minute is short-horizon.

As for complex tasks — at least in manipulation, Unitree focuses more on local motion and such. There are unsolved problems in manipulation. For example, when a robotic hand makes contact with the external world, it produces very complex physical phenomena and physical models. Then, under multimodal, high-dimensional visual input, how to complete relatively dexterous tasks while achieving very high success rates.

This has been the most critical challenge in manipulation for 50 years, and we're now attempting some work in this area.

Question 14: Manipulation — the robot manipulation problem — is also receiving very high attention right now.

Jianlan Luo: If robots truly achieve manipulation, that's AGI. It's intelligence more advanced than LLMs. If human civilization is zero to ten, LLMs at most rate a three. But if manipulation is achieved, it would be at least seven or eight.

Question 15: In the pursuit of robot AGI, what interests you most? Jianlan Luo: How to give this system stronger autonomous learning capability and generalization ability. In 2016, after Google published the first deep robot learning paper, not a single learning-based robot was actually deployed to the real world. But now things are different. The embodied intelligence research center we've newly established at AgiBot is neither a pure research institution nor a pure engineering deployment institution. It's an intermediate state, hoping to bridge the chain from basic science to technology deployment. Question 16: Influenced by large models, reinforcement learning is also becoming trendy in embodied intelligence. Jianlan Luo: Everyone is now looking in this direction, because we have DeepSeek R1, we have GPT-o1. In robotics, a field with 50 years of history, while many professors have done pioneering work solving control stability and other problems, my observation over the past decade is that progress in this field has always come from other fields — computer vision or NLP. Now there are several waves of people doing embodied intelligence: some from CV, some from learning-based approaches, some from core robotics — each has different perspectives. Question 17: Now many large companies, industrial players, including consumer electronics companies are entering embodied intelligence. What unique advantages do startups like AgiBot have? Jianlan Luo: Many players entering is actually a positive signal — it means growing attention. As the next-generation intelligent terminal, robots are naturally on consumer electronics companies' radar. They have very strong accumulated expertise in user experience, productization, cost control, and supply chain integration. Teams like AgiBot's advantage lies more in understanding the industry's underlying logic. They may be more vertical, more refined; we may be stronger in intelligence. Ultimately, the two directions will converge. Question 18: What cycle do you think embodied intelligence is currently in, and is it still a good time to enter? Jianlan Luo: Looking from 2016, I think embodied intelligence has gone through roughly a decade of exploration; initially it was called robot learning. I think now is a very exciting time. Within a few years, we'll see some successes in specific scenarios. Actually, worldwide right now, 5 million robots are deployed in the real world, but they're all blind robots — operating through absolute positioning, doing repetitive programmed work. As intelligence improves, we've entered the application window for robots. While those idealized, all-capable robots may take ten years or longer to arrive, robots with usage value in specific scenarios and continuous learning capability will come sooner. So now is the best time to enter and to break through.

From VLA to ViLLA: AgiBot Releases First General-Purpose Embodied Foundation Model GO-1 | BlueRun Ventures Portfolio Headlines Zhihuijun: The Wild Iron Man Begins a New Adventure | What's Blue About Entrepreneurship BlueRun Ventures Dialogues with Galaxy Universal, AgiBot, and Lingchu Intelligence: When Embodied Intelligence Enters Reality | The Road to AGI