XPENG Robotics, what level is it? After comparing with global robots, we found the answer.

Yesterday, XPeng Motors unveiled its new-generation robot, IRON. Due to its striking resemblance to a human, it made headlines in the tech circle.

Upon closer inspection, its body shape, walking posture, ear shape, etc., truly bear a remarkable resemblance to a human. If it weren't for the founder of XPeng Motors releasing a single-shot video, the suspicion of "human impersonation" might have persisted for a long time.

Of course, appearance is important. But as tech enthusiasts, we should focus more on its intrinsic value: What level is the XPeng robot at? Where exactly lies its scarce value? This article will elaborate on it from four aspects.

01

Flexibility and Precision

First, its physical hardware and design are relatively close to human flexibility and execution precision.

It has 82 degrees of freedom in its body and 22 degrees of freedom in its hands. How can we understand these data? You know Tesla Optimus (Tesla's humanoid robot), right? According to some sources, its prototype in 2022 had approximately no more than 40 degrees of freedom in the body and about 22 degrees of freedom in the hands.

Of course, after all, that was a machine from three years ago. Let's compare it with the latest domestic robots.

According to PencilNews' statistics, among the currently publicly released robots in China, most have between 20 - 50 degrees of freedom in the body, and few have more than 22 degrees of freedom in the hands.

Some people may ask: What's the use of pursuing these parameters? The higher the values, the stronger the ability for fine operations.

With 22 degrees of freedom in the hands, which is comparable to the flexibility of human joints, it can grasp, pinch, twist, write, plug and unplug, and perform fine assembly - it can do most of the work that humans do.

To help you better understand, we've prepared a set of charts.

It's important to emphasize here that although the flexibility of the XPeng robot is excellent, it's still a long way from "working for humans on a large scale." Why? No matter how flexible your body is, you also need a good "brain."

What we're actually referring to here is the "intelligence level of the robot's brain." The greatest characteristic of humans is the ability to use tools. If a robot has a poor brain, it's still just a tool, not a "person."

For example, can the robot understand human language? At this level, large models are probably the highest scorers.

Secondly, after it understands the language, can it understand the task well? For instance, if you ask it to clean the house, will it plan the steps of "tidying up the table first → sweeping the floor next → mopping the floor last"?

Thirdly, how is its real-time decision-making ability? When walking, if a ball suddenly rolls in front of it, it needs to be able to predict that "a child may follow" and immediately stop and wait. This might be a bit difficult for current AI.

Finally, can it perform stably? Completing an action once occasionally and completing it stably over the long term are two different concepts. This requires a large amount of data and algorithm iterations.

02

Capable of Screwing

The second valuable aspect of the XPeng robot is its ability to work in factories screwing, leading in the commercialization process.

The IRON robot has been integrated into the production line of XPeng's Guangzhou factory to assemble components of the "XPeng P7+ models." It can install 500 screws per day, saving over 300,000 yuan in labor costs annually.

How significant is this breakthrough? Based on PencilNews' statistics of global cases, it's considered relatively advanced.

In Japan, a metal processing company, Fuji Sengen Kogyo Co., Ltd., has been using the duAro robot from Kawasaki Heavy Industries to screw. This robot can carry a load of 3 kg and has a repeat positioning accuracy of ±0.05 mm.

Kawasaki's DuAro (Collaborative Robot)

Of course, this robot is not a humanoid robot but a dual-arm collaborative robot.

In China, UBTECH's industrial humanoid robot, Walker S1, has also started working in factories screwing. On the electric vehicle production line, Walker S1 has begun training and deployment for component assembly, screwing, parts sorting, and logistics handling.

Additionally, a South Korean media report pointed out that in Jiangsu, China, a company and Galbot have developed a "humanoid robot" for checking whether screws are tightened and whether there are cable installation issues.

From these cases, we can see that although there are precedents for robots "entering factories" globally, some are not in humanoid form, some are only in the training and deployment stage, and some are only for inspection work.

More importantly, the XPeng robot is not only used for screwing but has also reached a cooperation with Baosteel for industrial inspection and other scenarios. This shows that the commercialization of the XPeng robot is relatively advanced.

03

Mass Production Stage

The third valuable aspect of the XPeng robot is mass production. Of course, this only means it's relatively advanced, not the only one.

The mass production of IRON is scheduled for the end of 2026. Currently, there are approximately hundreds of humanoid robots deployed on XPeng's production line.

Let's compare and see what level it is globally.

First, look at Tesla (Optimus). Elon Musk has publicly stated multiple times that there will be limited mass production between 2025 - 2026, and he hopes to start large-scale production in 2026.

Secondly, there's Figure (a Silicon Valley company). It has launched the Figure 02 series in factories such as BMW for pilot projects and was regarded as one of the few commercialization test cases that achieved "factory implementation" between 2024 - 2025.

Thirdly, there's Apptronik (a company in Texas, USA). It's conducting a cooperation test with Mercedes - Benz and has publicly stated that its goal is to start manufacturing commercial units and expand production capacity in 2026.

In this regard, if XPeng can really achieve mass production in 2026, it should be able to enter the global first - tier.

But after all, this is just a plan, and mass production is far from that simple.

Take "consistency" as an example.

An industrial - grade robot may have more than 1,200 components. If it's to be mass - produced, it means the consistency error of the motors and drivers of these components must be ≤ ±1%.

This means that each joint needs to have stable force control, stable position control, stable temperature compensation, and stable communication delay; otherwise, it will shake, drift, and be inaccurate. So mass - producing humanoid robots is very difficult.

Although it's difficult, it's also very important. Why? If humanoid robots aren't mass - produced, the price won't come down.

Figure, Apptronik, and Tesla have all publicly mentioned that mass production is the only way to reduce the overall cost from six - digit dollars to five - digit or even four - digit dollars. Otherwise, it's difficult for robots to demonstrate their "cost - effectiveness" advantage, and customers won't have the motivation to purchase.

04

Self - developed Chips

The fourth valuable aspect of the XPeng robot may be the self - controllability of core hardware.

It's equipped with (partially) self - developed chips - the Turing AI chip, with a computing power of 2,250 TOPS, three times that of current mainstream chips.

Many people may not understand: Why develop chips independently?

In humanoid robots, the chip is equivalent to the cerebral cortex. How powerful its performance is determines the upper limit of the brain's intelligence. If such an important thing relies on external sources (NVIDIA, etc.), it may mean:

1. The computing power architecture is determined by others, and you have to wait for their support to adapt to new algorithms.

2. The cost and supply cycle are restricted.

3. The algorithm optimization space is limited (for example, energy consumption, delay, and power consumption optimization all depend on the support of the chip manufacturer).

Additionally, XPeng has multiple products that need chips, such as cars, aircraft, and robots. These products can share an intelligent base: Cars need autonomous driving, flying cars need real - time navigation and stable control, and needless to say about robots.

So, what's the technical content of the self - developed chip by XPeng?

It mainly lies in the support for high computing power and large models. For example, the Turing chip is equipped with a 40 - core processor and claims to support a large model with 30 billion parameters.

These parameters seem very strong, but how it performs in robots and complex scenarios still needs time to verify. It's just that XPeng's courage and innovative spirit in independently developing chips are worthy of recognition right now.

This article does not constitute any investment advice.

This article is from the WeChat official account "PencilNews" (ID: pencilnews), author: Xi Wen. Republished by 36Kr with permission.