The AI capabilities of embodied intelligence are still at the GPT-3 era. It is expected that multiple companies with a market value of over 100 billion will emerge | Recap of OpenTalk

Text | Liu Jingqiong

Editor | A Zhi

On June 25, 2025, 36Kr jointly organized an OpenTalk live event titled "Equipping AI with a Body: When Will Embodied Intelligence Usher in Its 'ChatGPT Moment'" with Yundao Capital. The event specially invited Fang Zhenghao, Managing Partner of Xiaomiao Langcheng; Lü Liyun, Founder and CEO of Yaole Technology; and Jin Ge, Founder and CEO of Lingyu Intelligence. The three guests conducted in - depth sharing on various aspects such as the investment logic of embodied intelligence and AI, the key infrastructure of embodied intelligence - flexible tactile sensors, and the progressive development path of embodied intelligence from L0 to L4, and interacted with the audience online. The following is a summary of the key points from this live event. Welcome to read, share, and collect it.

Investment Logic of Embodied Intelligence and AI

Guest Speaker: Fang Zhenghao | Managing Partner of Xiaomiao Langcheng, Bachelor of Science from Fudan University, EMBA from China Europe International Business School. Responsible for the company's investment, research, and channel development system construction. He has previously worked in overseas hedge funds, assisting in managing the assets of overseas institutional investors such as the Bill Gates family fund, with over 10 years of investment research experience. The investment cases he has led and participated in include: Shipu Testing (301228.SZ), Mufan Power, Aochuang Photonics, Yueqian Engine, Qiongche Intelligence, Songyan Power, West Lake Robot, Guangbenwei, ZStack Information Technology, Saizhuo Electronics, etc.

Keywords for Sharing: #Embodied Intelligence #Humanoid Robot #Artificial Intelligence

Fang Zhenghao believes that the AI capabilities of embodied intelligence are currently at the GPT3 stage, and the issues of data collection and pre - training have not been completed. However, with the improvement of the general and generalization capabilities of large models, there will be significant investment opportunities in both the artificial intelligence and embodied intelligence industrial chains, specifically manifested as follows:

• On the artificial intelligence side, at the bottom - level hardware computing power layer, the opportunities for startups in traditional computing power fields such as GPUs are limited. Institutions can pay more attention to new opportunities in next - generation new - type computing power and edge - side computing power. At the model layer, attention can be paid to some computing power integration opportunities that will arise after the mixing of heterogeneous computing power, including some opportunities for data - related annotation and services. At the large - model layer, attention can be paid to the integration opportunities of multi - modality and cross - modality. In particular, 3D spatial intelligence, as the foundation for AGI perception, may reconstruct the paradigm of human - machine interaction, so 3D model technology needs to be paid attention to in advance. Between the application layers, attention can be paid to niche opportunities such as heterogeneous computing power scheduling and inference acceleration. However, it should be noted that traditional cloud providers may squeeze the living space of the middle layer with their full - stack service coverage. At the terminal application layer, on the C - end, attention can be paid to general entertainment scenario applications that capture users' emotional value and start from high - frequency interaction needs. On the B - end, attention can be paid to applications that can improve efficiency in vertical industries (such as law, medical, etc.).

• On the embodied intelligence side, the biggest core barrier and threshold lies in the software layer. Firstly, in the robot's brain, which is currently the biggest bottleneck for all robots, attention can be paid to startup teams with core capabilities at the brain layer. On the other hand, at the cerebellum layer, also known as the lower - limb movement layer, teams with advantages in reinforcement learning and control algorithms have demonstrated some running, jumping, and flipping abilities of robots in the past one or two years, which is very helpful for the implementation and commercial promotion in certain scenarios. At the hardware layer, the supply chain of embodied intelligence is relatively short, and it is difficult for startups to build core barriers through unique hardware technologies and differentiation.

Investment Logic of the AI Industrial Chain and the Invested Cases of Xiaomiao Langcheng

Key Infrastructure of Embodied Intelligence - Flexible Tactile Sensors

Guest Speaker: Lü Liyun | Founder and CEO of Yaole Technology, former Chief Architecture Engineer of the Global Innovation Department of Harman, an international automotive electronics giant. She led the construction of a multi - modality sensor fusion computing platform, serving more than a dozen first - tier automobile manufacturers such as Maserati and Porsche. She is one of the earliest domestic participants in intelligent driving field projects and a senior expert in the early research and development of cloud computing platform architecture in China, deeply involved in the fields of intelligent driving sensor data processing and robot perception systems.

Keywords for Sharing: #Intelligent Perception #Flexible Tactile Sensor #Embodied Intelligence #Embodied Perception

Lü Liyun introduced several common tactile sensor solutions on the market and their respective characteristics:

• Hall - type sensors, being at the chip level, are a good solution for the hand area of robots. They have high perception accuracy but also higher costs.

• Capacitive sensors have relatively high sensitivity, and many capacitors can be made into film - type products. Their thickness and sensitivity are also suitable for small hand areas. However, their disadvantages are poor stability and durability.

• Piezoelectric sensors are mainly used to detect dynamic forces and are relatively sensitive. However, they are easily affected by the thermal response effect, and the temperature zone stability control of the application scenario is relatively strict.

• Piezoresistive sensors have a sandwich structure, that is, upper and lower electrodes plus a middle pressure - sensitive layer. Their characteristics are relatively stable, good durability, and they are suitable for large - area durable application scenarios. However, they are easily interfered with in complex electromagnetic environments.

Yaole Technology has improved the materials and processes on the basis of the original piezoresistive solution. Using the original metal yarn integrated weaving technology, it has produced fabric pressure sensors. Both the electrodes and the pressure - sensitive layer of this sensor use fiber - grade materials. As the electrodes of the sensor, metal yarn has the greatest characteristics of low resistance, low elasticity, and high strength compared with metal - coated fibers and conductive materials, which can solve the problems of durability and stability of tactile sensors under other technical routes. In addition, the fabric - type sensor pressure - sensitive material independently developed by Yaole mainly solves the problems of conductivity uniformity and durability, and its performance has passed the automotive - grade test. Previously, many conductive materials on the market were not used as sensors and could only achieve anti - static and conductive functions, lacking uniformity and durability.

For a long time in the future, Lü Liyun believes that fabric - type and traditional electronic - printed sensors will coexist. The industrial chain of printed sensors is mature, and they have a large application space in some small - area tactile scenarios. However, due to printing technology or material limitations, they are prone to falling off due to friction, cleaning, etc. In large - area tactile scenarios, especially those with high requirements for flexible adaptation to irregular surfaces, such as intelligent car seats and robot electronic skins, fabric - type tactile sensors are still a better choice.

Comparison of Fabric - Type Sensors with Different Technical Routes

Discussion on the Progressive Development Path of Embodied Intelligence from L0 to L4

Guest Speaker: Jin Ge | Founder and CEO of Lingyu Intelligence, undergraduate from the Department of Automation at Tsinghua University, MBA from the School of Economics and Management at Tsinghua University. He was formerly the Managing Partner of Yuanjing Venture Capital and the Vice President of Aoliang Photonics. He has many years of successful investment and entrepreneurship experience in the high - tech field and has invested in and incubated several early - stage hard - tech enterprises. Lingyu Intelligence is a new - generation embodied intelligence company that mainly focuses on human - machine hybrid intelligence. Founded by a top - notch motion control team from the Department of Automation at Tsinghua University, its mission is to "create a benchmark for the practical application of embodied intelligence and liberate humans from 'dangerous, heavy, and boring' work".

Keywords for Sharing: #Remote Operation #Human - Machine Hybrid Intelligence #MAAS #Arm - Hand Integrated Control

Jin Ge observed that currently there is an impossible triangle in embodied intelligence, that is, it is difficult to balance generality, performance, and autonomy under the current and even the technological level in the next 3 - 5 years. Generality means that the robot is not dedicated to a specific scenario but can perform many different tasks. Performance refers to, firstly, reliability, that is, how high the success rate of the robot in performing a task is, and secondly, efficiency, that is, when the robot and a human perform the same task, how much faster or slower the robot is than a human. Autonomy means whether the work requires human intervention or can be completed by the robot itself.

Subsequently, Jin Ge introduced two common methods for improving the autonomy of robots:

• The first is for enterprises to try to reach L4 directly, increasing the operation success rate to over 99.9% and achieving the fully autonomous working state of robots in multiple scenarios. However, this path takes a long time and incurs high costs. Currently, robot data is extremely scarce, and enterprises need a large amount of real - machine data, resources, and funds to train a high - intelligence AGI robot that meets human expectations.

• The second is to refer to the current progressive thinking of autonomous driving, gradually improving from L0 to L2 and then to L4. That is, first put the robot into commercial use, continuously collect interactive data, and then gradually upgrade the robot's intelligent system. The biggest advantage of this method is that enterprises can make up for the short - board in data and obtain revenue earlier.

Following the second idea, Jin Ge believes that a more economically feasible method at present is to establish a MAAS (Manipulation AS a Service) platform. That is, in daily simple scenarios, the robot operates autonomously. When encountering complex and dangerous situations, the robot will call a real person or a "human - like" model in the cloud. The real person or the cloud model will take over the robot through remote operation to complete the next step of operation. This method can achieve one - to - many takeover of robots. While improving the autonomy of robots, it can also better meet the personalized needs of users.

Progressive Development Path of Embodied Intelligence from L0 to L4

What Are People Discussing About Embodied Intelligence?

We selected some representative questions from the live - interaction session and the guests' answers, which are presented after editing:

Q1: Are the valuations of some current embodied intelligence projects too high? How do you view the bubble phenomenon in the current track?

Fang Zhenghao: Bubbles are an inevitable part of the development process of the technology industry. Looking ahead ten years, all developed countries in the world and China will face a huge labor shortage. The population engaged in white - collar and blue - collar jobs will decline sharply. Ten years later, the annual employment cost for these positions in China may exceed 30,000 US dollars, while in developed countries, it will rise to 50,000 - 80,000 US dollars. With a global labor shortage of hundreds of millions and an employment cost of 30,000 - 50,000 US dollars, there will be an overall market of 2 trillion US dollars, and this labor market will be partially or fully replaced by robots. From this perspective, embodied intelligence will at least be a trillion - dollar industry in the future, and there will definitely be several companies with a market value of over 100 billion US dollars. Looking at the end - goal, even if there is a certain bubble in this track and some startups have high valuations, I still think it has a certain degree of rationality.

Q2: Will embodied intelligence ultimately evolve into general - purpose humanoid intelligent agents or robots dedicated to vertical scenarios?

Fang Zhenghao: These two will coexist in the future. In terms of usage volume, in the long run, humanoid robots will have the largest volume. From the perspective of first - principles, we design humanoid robots based on human living and working habits, so theoretically, their versatility is the highest. However, in most current vertical scenarios, dedicated robots have lower costs, higher reliability, and lower energy consumption. They may not need to be in a quadruped, biped, or even humanoid form to meet the usage requirements. Therefore, dedicated robots for vertical scenarios will definitely have their own living space for a long time.

Q3: What are the advantages for Chinese embodied intelligence enterprises to go global?