Does AGI Need a Body? Embodied Intelligence with 它石智航's 陈亦伦
Summary
陈一轮 is betting not on “yet another robotics company,” but on robots converging from fragmented tools into a general-purpose platform. He says the signs were already visible in 2022, and he founded the company with a partner on February 5, 2025, aiming to make embodied intelligence a market an order of magnitude larger than autonomous driving; yet while capital and talent continue to pour in, the commercial “exit” remains narrow.
The data scale required for an embodied foundation model may be 10 million hours, not a more elegant network architecture. Autonomous driving requires roughly 1 million hours of refined data to reach a product-grade release; robots, with 10x the complexity, would therefore require at least another order of magnitude. Teleoperation is real and effective, but its collection speed, cost and limited scene coverage make it better suited to training vertical small models; as a data-acquisition route to an embodied foundation model, it is a “dead end.”
The point of both end-to-end and full-stack is to shorten the AI loop, not to own every component. The team develops its own data-collection systems and core execution while working to narrow the gap between collection and execution, but buys cameras, chips, motors and gears; its operating principle is that “data is the source code, and the model is only the compiled result,” which requires tight coupling among collection, training, deployment and failure feedback.
Human data will not permanently cap robots at human capability. Imitation learning first has the “teacher” provide the steps to solve a problem, while reinforcement learning turns the teacher into a judge of right and wrong, giving the student a chance to surpass the teacher; the most valuable demonstrations may not be standard movements, but “the least deliberate little actions,” such as edging, probing and retrying.
Vision and touch are the 2 necessary modalities 陈一轮 identifies, and touch determines whether a task can clear the final centimeter. Autonomous driving is a non-contact system, while robots operate through contact, so failures often occur in the final centimeter; usable sensors must be “rich and accurate,” with the current focus on fingertip or palm-tip tactile sensors.
Wiring harnesses are not the most impressive demo, but they are the team’s first million-scale application. Soft, precise manipulation generates large amounts of real data while subjecting the system to industrial tests of takt time, reliability and quality; the team has worked on it for about 14 months and, as of then, had begun small-batch industrialization, with 陈一轮 judging that another roughly 14 months could “solve it completely.”
陈一轮 expects China to win the embodied-intelligence race, not because of a lead in any single model, but because it has a more complete data–scenario–hardware–AI chain. The US is stronger in compute, AI talent and its startup environment, but lacks some industrial scenarios and data-collection systems; US teams may pin their hopes on YouTube data, but the shortcut to starting the flywheel is still “real, effective data.”
On AGI, 陈一轮 believes Physical AI must interact with the world through sensing and action to change it; without a body, that interaction is impossible. Coding moved first because a sandbox allows interaction, verification and massive trial and error. The host therefore summarized AGI as potentially requiring action and a body, while 陈一轮’s end-state formulation is: “Action Is All You Need.”
Deep dive
1. Robotics must converge from fragmented tools into a general-purpose platform to become a major market
After a decade in autonomous driving, 2 years teaching at Tsinghua, and a return to entrepreneurship, 陈一轮 was trying to answer a question more fundamental than whether models had improved: are robots still a collection of specialized devices, or has a general-purpose technical and commercial path emerged?
The signs appeared in 2022, and the conditions for starting a company gradually came together. On February 5, 2025, he founded a company with a partner to bring embodied intelligence into everyday work and life and turn it into a market an order of magnitude larger than autonomous driving.
The experience also changed his view of innovation: results do not require first coming up with an “especially remarkable method.” Important outcomes force a team to keep searching for lemmas and correcting its path. “It is a process of exploration” that requires “starting with the end in mind,” rather than betting on a single technical masterstroke.
2. Data distribution determines model architecture; the bubble comes from hot methods and cold commercial outlets
陈一轮 draws a sharp contrast between industry and academia. In companies, discontinuities in data volume and distribution can create a generational product lead, while differences in network architecture may not matter much when the data generation is the same. Universities, by contrast, tend to equalize the data and compare methods on a common benchmark. “The distribution and type of data determine the network architecture.”
He does not deny that embodied intelligence has a bubble, but attributes it to a missing objective: resources are concentrated on methods, supply and technology, while too few people ask what problem is being solved and what value is being created. The industry’s capability frontier is rising visibly each year—sometimes advancing more in 1 year than in all the years before—but capital and talent continue to enter while genuine commercial outlets remain narrow.
3. Autonomous driving could gradually shift to AI; general-purpose robots must be AI-native from birth
From an AI perspective, 陈一轮 says the 2 fields are “exactly the same.” The difference is that autonomous driving could initially be advanced without AI: the problem could be modeled in the relatively closed world of people, cars, roads and traffic lights. What ultimately forced the shift to AI was the dynamism, uncertainty and interaction created by mixed human–vehicle traffic.
The host asked whether the more open and complex robotics environment was simply an extension of autonomous driving. 陈一轮’s answer was stronger: general-purpose robots do not have the blessing of shipping before AI is ready. They therefore cannot repeat the gradual move from rule-based systems to AI; they must start with an AI-native approach.
That also means the old organizational separation of perception, planning and action will not simply reappear. The problem must be defined end to end, then kept in a continuous loop of data, deployment feedback and failure samples.
4. 10 million hours is the foundation-model threshold; teleoperation can feed only vertical small models
陈一轮 estimates the data requirement by working backward from autonomous driving. A system needs roughly 1 million hours of carefully compressed, high-quality data to reach a product-grade release. If embodied-model capabilities are at least 10x more complex, an embodied foundation model may require 10 million hours of distilled data. The host noted that even leading companies currently have only a few hundred thousand hours.
He still believes 10 million hours “may be enough.” Moving from 10 million to 100 million hours would continue to improve performance, but marginal returns would begin to decline; the biggest scaling factor would shift to a second scaling curve rather than simply adding more data.
Teleoperation’s advantage is that the data is real and works. Its drawbacks are slow, expensive collection and insufficient penetration across scenarios. 陈一轮 therefore sees it as a solution for small models trained on vertical tasks. If the goal is a true embodied foundation model, this data-acquisition route “will never happen”—it is a “dead end.”
The scalable path is to record the work and daily activities people already perform, just as the internet naturally accumulates text and people generate driving data simply by driving every day. For robots, the ideal is: “You work, you live; I record you,” provided the collection system is sufficiently user-friendly.
5. Human demonstrations provide the prior; action is the prediction target in the physical world
In response to concerns that human data might lock robots to the human ceiling, 陈一轮 distinguishes 2 stages. Imitation learning asks the student to copy the teacher’s solution steps and quickly close the capability gap. Reinforcement learning then demotes the teacher to a judge of right and wrong and of the boundaries, at which point the student “has every chance of surpassing the teacher.”
He compares current robots with early large language models: they “have not yet formed a species,” but already show strong learning ability. AI does not develop by becoming a generalist before becoming a specialist. It first specializes rapidly in one area, expands into specialization across multiple areas, and finally lifts the shared baseline so those capabilities connect.
The robotics equivalent of the “next token” need not be a discrete token. A language model treats context and text as state, predicts the next state through next-token prediction, and uses compilation or testing as the reward. Physical AI data consists of continuous sensor and action streams, but the underlying structure is still state–action–reward: how the world is changed by action.
To 陈一轮, World Model is a neutral and ancient term; the important question is what it contains. Before founding the company, the team once called it AI World Engine because it predicted that World Model would be “used to death” 2 years later. Methods exist to solve problems, and once they do, they eventually become baselines.
6. “The least deliberate little actions” condense skills humans have already mastered
陈一轮 uses inserting a key into a lock as an example. A human does not align it with precision and push it straight in like a traditional robot. They casually place it near the opening, search along the edge, and try again if it does not work. These probing movements materially improve task robustness and carry high learning value.
His explanation is a “cerebellar response.” Such movements do not exist when a skill is first being learned; after enough practice, countless tiny, unconscious adjustments emerge. Capturing them is equivalent to distilling skills humans have already internalized. “Our least deliberate little actions are the most valuable.”
7. Full-stack does not mean building everything; it means controlling the AI loop
陈一轮 defines the most important early-stage capability as rapid iteration. Data collection, sensing, training, execution and feedback are beads strung together by an AI loop. If a critical bead sits outside the company, iteration repeatedly encounters friction; the parts that have not yet been standardized are usually where the value lies.
The team therefore controls its data-collection methods, AI input devices such as gloves, and the core execution layer, with the aim of narrowing the gap between captured actions and robot execution. It does not build cameras, chips, structural components, motors or gears; the team designs the joints. Standardized components can remain with industrial partners.
陈一轮 uses software engineering as the analogy: “Data is the source code; the model is only the compiled result.” AI engineers should worry about data first. The company is building a large “compilation system,” not treating model training as an isolated department.
He recalls managing an autonomous-driving team of roughly 1,500 people, including about 200 in perception and 300 in planning. Traditional software relies on people and organizations to transmit information and decompose problems, eventually hitting a ceiling as headcount grows. AI is an “automated mechanism for propagating and reducing problems,” giving a small team a chance to complete a large system.
8. Touch determines the final centimeter; contact sensing is essential beyond vision
陈一轮 identifies only 2 necessary modalities: vision and touch. He is a “staunch advocate of touch” because autonomous driving is a non-contact system while robots are contact systems; tasks often fail not at a distance, but in the final centimeter.
He explains vision–touch fusion through 2 world models: one for the ordinary visual world, and another formed by blind people through touch. A blind person can also tell whether an object is round or sharp; the 2 forms of representation ultimately converge in the same frontal lobe.
A good sensor must be “rich and accurate.” He cites millimeter-wave radar as a counterexample because it is sparse and noisy, while acknowledging that tactile sensing has not yet reached the ideal state. The current focus is on fingertip or palm-tip sensors, and he says tactile sensors are “essentially visual sensors.”
9. Wiring harnesses are a million-scale application; demanding industrial metrics are unavoidable
陈一轮 compares Physical AGI with the Goldbach conjecture. The destination can command broad agreement; what separates companies is which lemmas they choose to reach it. The team does not want to serve scattered requirements indefinitely, so it is starting with concrete applications that have a scale of at least 1 million.
It identified about 3 such scenarios, using China’s roughly 2 million Robotaxi vehicles as a reference point. Wiring harnesses are the best “first appetizer”; the other 2 scenarios have not been disclosed. The work is not sexy to ordinary viewers, but it is a persistent, difficult and high-value headache for users.
Wiring harnesses involve extensive soft, precise manipulation. A manufacturing site can generate large amounts of data quickly while offering a relatively in-house environment. The trade-off is that the system must meet strict takt-time, reliability and quality requirements simultaneously. 陈一轮 does not see those constraints as a detour, but as a necessary stage on the way to real-world robotics.
The team has spent about 14 months on the problem and has begun small-batch industrialization. 陈一轮’s clear judgment is: “Give me roughly another 14 months, and this problem can be completely solved.” The premise is that the methodology is correct.
10. China’s edge is a real-data flywheel; AGI’s end state is a body and action
陈一轮 sees China and the US competing out of alignment. The US has not yet produced a benchmark project in embodied intelligence for China to follow, so China has a field of many competing approaches. The US is stronger in compute, AI talent and a tolerant startup environment; China is more active across scenarios, sensor data collection, hardware and the industrial chain.
The embodied-AI loop is longer than the language-model loop, and the US still has structural gaps in some industrial scenarios and data-collection systems. He therefore concludes that “the China team will still win.” US teams may pin their hopes on YouTube data or other sources, but the shortcut to starting the AI flywheel is often still “real, effective data.”
The host cited a possibly 2016-era optimistic forecast for autonomous driving in 2020 and asked whether 2030 could become another “magic number” for robots. 陈一轮 first pushed back on the idea that autonomous driving failed to reach deployment, then said that if robots are not everywhere, the reason will be the commercial outlet and whether the problem has truly been solved—not that technology will stop iterating rapidly over the next 5 years.
On AGI, 陈一轮 notes that Dario and Ilya more often discuss Super Intelligence than General Intelligence. AI will first surpass humans at high-impact points and then connect those capabilities into a whole, like “a country of genius in a data center.” Digital models change text or tokens; Physical AI changes the state of the world, so its inputs and outputs must ultimately be sensing and action.
Coding became the “appetizer” because a sandbox allows interaction, verification and large-scale trial and error. 陈一轮 believes real-world intelligence must obtain the same loop through a body. Without a body, there is no interaction with the world and no way to keep generating data in the second half of AI. The host therefore summarized AGI as potentially requiring action and a body, while 陈一轮’s end-state formulation is: “Action Is All You Need.”