Embodied intelligence didn't materialize in its inaugural year.

In early December, two leading domestic universities launched majors in embodied intelligence one after another. This seems like a grand handshake between the industry and academia, making the audience fantasize about a bright future for domestic embodied intelligence.

On December 8th, Zhipu Robotics announced the mass production of 5,000 units of its products. The year 2025 was a year of rapid development for domestic embodied intelligence. From taking the stage at the Spring Festival Gala at the beginning of the year to entering top - tier academic halls at the end of the year. There are more and more participants in the embodied intelligence track, and people's understanding of it is also becoming richer and more diverse.

Some people think that embodied intelligence is the physical representation of AI, while others believe that it is a new interaction paradigm. It's difficult to find a definition that everyone agrees on, but there is one view that everyone shares: Embodied intelligence is relevant to me.

What can't embodied intelligence do?

In the era of large models, all industrial systems are worth reinventing.

Optimists are asking, "What can't embodied intelligence do?"

Onlookers are watching, "Embodied intelligence can fold clothes, handle logistics, play football, attract people, and also..."

There is still a huge gap between the prosperous demos and the actual implementation. For the sake of financing, embodied intelligence projects burn through their "imagination". A demo can showcase a certain ability, but different demos may not necessarily demonstrate different algorithmic capabilities. Currently, there is a phenomenon of "horizontal expansion" of demos in the industry, that is, all the "tricks" are actually the same, using permutations and combinations of the same elements to create a false sense of prosperity. This contradiction is the difficulty in balancing the expansion of embodied intelligence and the accumulation of technology; too deep a technological accumulation will result in high investment, difficulty in achieving results, and ultimately being eliminated.

Demos are everywhere, and December has arrived, but embodied intelligence has not been implemented yet. The first reason is that the capabilities of embodied intelligence have not reached the threshold for large - scale implementation. When discussing the capabilities of embodied intelligence, we need to look at the success rate, rhythm, cost, and reliability of simple tasks.

Although some embodied intelligence systems perform stably in highly structured laboratory environments, with a task success rate possibly exceeding 80%, the success rate may drop significantly in real - world environments. More importantly, even if the success rate of a single task can be improved, for long - term tasks, the overall success rate is the product of individual success rates, and multiplying numbers less than 1 will naturally result in a lower outcome.

The actual implementation of embodied intelligence requires finding suitable scenarios and ensuring an increase in the success rate. It's not yet known how long this path will take, but one thing is certain: the growth of the scale of embodied intelligence is not absolutely positively correlated with the growth of demos. Just like the story of "The Boy Who Cried Wolf", people heard in 2015 that "this is the year of embodied intelligence".

The ChatGPT moment of embodied intelligence

The ChatGPT moment for embodied intelligence will occur when it is applied in the real world and everyone wants to use it.

Capital is eager to explore the application scenarios of embodied intelligence, but companies in the field of embodied intelligence still need to think about how to apply it and in which scenarios. On December 3rd, Tesla showed a video of a robot running. Yes, the embodied intelligence can run, but then what? In what scenarios would we need a robot to replace humans in running?

The industry currently attributes the future implementation directions of embodied intelligence mainly to three scenarios: commercial service scenarios, industrial scenarios, and household scenarios. It is highly likely that the implementation order of embodied intelligence will be commercial services first, then industrial services, and finally enter households.

This order is mainly because embodied intelligence requires a large amount of data training to build a world model. In this model, embodied intelligence should be able to think and predict the next step. However, it's like a chicken - and - egg problem. Since embodied intelligence doesn't have the opportunity to collect a large amount of data in real - world scenarios, it's unable to quickly build a model. Commercial scenarios, especially hotels, are easier to train because of their relatively fixed environments. In terms of value, food - delivery robots have indeed reduced labor costs.

Industrial scenarios have high requirements for efficiency, and the efficiency of replacing automation will be a hard threshold. We can see that even though robots can perfectly reproduce some industrial operations, they are not as fast as human hands. From the user's perspective, paying for slower "human resources" is a losing deal. From the technical perspective, due to the fragmentation of industrial scenarios and the difficulty of data collection, it's hard to increase the scale, let alone break through the cost and efficiency limitations.

Finally, let's talk about household services. There are two extreme views on the prospect of embodied intelligence entering households. If it only needs to play the role of companionship and conversation, perhaps embodied intelligence can quickly enter the consumer market. After all, from smart home devices to current AI toys, human - machine interaction is not new. However, if embodied intelligence is to truly become a "family member" in a household, it will face security and cost issues. Embodied intelligence defined as a "family member" often covers medical and elderly - care scenarios, and its security should be examined more carefully.

The path to the popularization of embodied intelligence may be from specialization to generalization. At the beginning, it may be the stable execution of single - task in single - scenario; then it will transition to the execution of multiple tasks in single - scenario; finally, it will achieve the stable execution of multiple tasks in multiple scenarios.

The development of embodied intelligence also requires industry consensus, that is, a set of benchmark tests. Sports competitions can't reveal the real differences in embodied intelligence. Breaking through this requires the combination of industry, academia, and research. In the academic field, in addition to Tsinghua University and Shanghai Jiao Tong University, which have announced the addition of majors in embodied intelligence, a number of domestic universities are also applying to offer such majors.

The prosperity and anxiety of embodied intelligence

For thousands of years, people have been dreaming of creating some artificial things that can automatically complete the work that can only be achieved by human wisdom and ability.

In Homer's "Iliad", Hephaestus, the god of blacksmithing and carving, created metal robots and servants made of gold to help him with miscellaneous tasks. Aristotle predicted the emergence of automated tools, making labor unnecessary. In "Liezi", it is described that the craftsman Yan Shi made a lifelike, singing, dancing, and even emotional "mechanical puppet" for King Mu of Zhou. In "Gulliver's Travels", a mechanical device is described. With it, "the most stupid and ignorant person can write books on philosophy, poetry, politics, law, mathematics, and theology without relying on talent or learning."

In the past, people's imagination of embodied intelligence was always about replacing humans in boring, repetitive, and low - value work; at the same time, people were worried that they would develop into "gods" that "control" humans. The impact of AI on the labor force in 2025 illustrates this point. Therefore, the future of embodied intelligence may not necessarily be to replace repetitive work. Perhaps it is more meaningful for them to replace humans in dangerous work.

Although there is anxiety about the future of embodied intelligence, it has brought prosperity to many industries. For the chip industry, a large number of chip manufacturers have found new growth spaces for their products.

On the edge side, many domestic chip manufacturers have released products for embodied intelligence. GohighTech released the G32R501 real - time control MCU, which can meet the high - computing - power, high - efficiency, and high - precision performance requirements of embodied robots in perception and decision - making, motion control, and efficient human - machine interaction. With the full - stack motor - specific chip of "MCU + Driver + IPM" as the core and paired with GohighTech's self - developed motor algorithm platform, it can be applied to core scenarios such as robot joints, industrial encoders, and frameless torque motors, constructing the "nerve center" of embodied intelligence.

The N32H7 series MCU of Nationz Technologies provides powerful computing power and real - time response capabilities with its multi - core heterogeneous architecture and ultra - high main frequency, meeting the strict requirements of humanoid robots for complex control and high synchronization. Its built - in CORDIC coprocessor can efficiently complete mathematical calculations such as trigonometric/coordinate transformations involved in kinematics, significantly reducing the CPU load.

The MR series robot chips of Allwinner Technology use a 12nm process, integrating a CPU + GPU + NPU heterogeneous architecture. With a computing power of 3 - 4 TOPs, a power consumption of only 5W, and support for millisecond - level response, they provide core computing power for motion control and environmental perception for products such as Xiaomi CyberDog and Unitree series. The cost is only one - third of that of NVIDIA Jetson Nano.

The RK3588 of Rockchip uses an octa - core 64 - bit ARM architecture, combining 4 high - performance Cortex - A76 cores (with a main frequency of up to 2.4GHz) and 4 energy - efficient Gortex - A55 cores (with a main frequency of 2.0GHz), showing excellent multi - task processing and complex calculation capabilities. It has a built - in NPU with a computing power of 6 TOPS, supporting multiple data types and mainstream deep - learning frameworks, and can efficiently process AI tasks such as image recognition and voice interaction. Industry insiders revealed that Rockchip has shipped tens of thousands of relevant products in the field of embodied intelligence.

Biwin Storage said that it has launched products such as eMMC, UFS, BGA SSD, LPDDR4X/5/5X for the field of embodied intelligence and is actively expanding its customer base among the leading players in this field. According to the disassembly report of a third - party media, Biwin Storage's LPDDR4X and eMMC storage products have been used in Unitree's Go2 intelligent robot dog.

On the computing - power side, Intel and NVIDIA are still the core players in providing "top - tier" computing power for robots. As mentioned before, the VLA of embodied intelligence needs to build a world model, and building the model will inevitably drive the demand for computing power. Intel meets the different load requirements of motion control and AI inference by launching a heterogeneous system of GPU + NPU + CPU, enabling the operation of the VLA model.

In addition to focusing on hardware, NVIDIA has launched the NVIDIA Cosmos platform to accelerate physical AI. This platform is an integrated platform that can integrate cutting - edge generative world foundation models (WFM), advanced tokenizers, guardrails, and efficient workflows for accelerating data processing and management. It provides support for world model training and accelerates the development of physical AI for autonomous vehicles (AV) and robots.

Since it involves aspects such as mechanical control and edge computing power, the chip suppliers for embodied intelligence highly overlap with those for automotive chips. The development of embodied intelligence also shows some similarities to that of the automotive industry.

In 1885, Karl Benz built the first tricycle powered by a gasoline internal combustion engine. Perhaps the "social role" of robots now is like that of cars between 1900 and 1910: they are technological wonders but not social infrastructure. Now, cars have become a common sight everywhere. The development of embodied intelligence will take some time, but not a century.

What is certain is that we can see that the role of embodied robots is not yet defined, but it's not due to a lack of ability.

This article is from the WeChat public account "Semiconductor Industry Insights", author: Liu Qian, published by 36Kr with authorization.