He Xiaopeng: Not all car manufacturers can make humanoid robots

The wave of “creating humans” in the automotive industry continues to rise, with nearly ten domestic automakers developing humanoid robots independently. However, Xiaopeng Motors Chairman and CEO He Xiaopeng, a National People’s Congress delegate, warns that not all car companies are suitable to enter the humanoid robot track.

In an interview with Southern Metropolis Daily and other media on March 7, he stated that many automotive companies have adopted an integrated R&D model, which involves integrating technologies from different suppliers to ultimately deliver a high-quality product to users. This is a successful business logic. But for embodied intelligence, only when hardware is self-developed, software is self-developed, hardware and software are cross-domain integrated, and commercial sales capabilities are established, can products truly be made well.

“The development difficulty of humanoid robots is dozens of times that of a car,” He Xiaopeng admitted. If a vehicle’s assisted or autonomous driving features are not satisfactory to users, then it’s likely that self-developed humanoid robots by these manufacturers will also struggle to gain user trust.

Is the timing for humanoid robots to enter factories not yet ripe?

At XPeng’s Technology Day on November 2025, the company’s self-developed IRON humanoid robot attracted attention with its model-like “catwalk.” To achieve mass production of humanoid robots by the end of 2026, XPeng Motors announced on February 25 that it will build a full-chain mass production base for humanoid robots at the Embodied Intelligence Industrial Park in Guangtang Innovation City, Tianhe District, Guangzhou.

He Xiaopeng believes that truly usable robots must first meet automotive-grade standards, or even exceed them. The reason is that a car has only one engine; if it fails, users will directly think the vehicle has a major problem. A robot, however, integrates dozens of “engines”—joints. From a probability perspective, the likelihood of failure is greatly increased.

Yao Maoqing, partner at Zhiyuan Robotics, previously expressed similar views to reporters: in all manufacturing stages, strict adherence to automotive-grade standards is necessary to ensure product stability and consistency. Poor stability could hinder customer production lines and cause significant economic losses.

Some humanoid robot companies prioritize factories as the first step for commercialization. He Xiaopeng believes that in the short term, this is not a wise choice in China; it is more suitable for countries with higher labor costs, such as Europe and the United States. XPeng’s mass-produced IRON humanoid robot will first be applied in commercial scenarios like guiding and shopping, with future applications also likely to be related to automobiles.

Since the end of 2025, discussions about product homogenization and “reshuffling” in the humanoid robot industry have increased. He Xiaopeng views homogenization as a normal phenomenon in the early industry stage. He predicts that from 2027 to 2029, global humanoid robots will show more differences in intelligence, appearance, and deployment directions. Once one or two benchmark humanoid robots emerge, other companies will evaluate whether they can keep up—those that can will continue competing, while those that cannot will be forced to transform, leading to an industry “reshuffle.”

To improve robot intelligence, He Xiaopeng calls for more focus on the robot’s “brain” and “cerebellum” capabilities, as well as data accumulation for embodied intelligence models. In a proposal for the Two Sessions, he mentioned that most humanoid robots in China currently rely on software rule control, demonstrating strong capabilities at the motion control system level, but lack advantages in the “brain” (autonomous thinking and decision-making) and “cerebellum” (motion control) coordination systems, scene task generalization, and commercialization prospects.

He Xiaopeng previously told media in January that software rule-driven robots have obvious limitations: “Robots sold to factories need to scan maps, specify objects to pick up and their weights, and move from point A to B based on rules. Slight changes in physical location require reprogramming, similar to industrial robotic arms lacking generalization.” A founder of a humanoid robot company recently told Southern Metropolis Daily that some humanoid robots currently used in factory POCs (proof of concept) are “not powered by AI but based on traditional rule-based methods.”

He Xiaopeng hopes that XPeng’s mass-produced robots will shift toward AI-driven paths: not pursuing extreme performance like flips or heavy lifting, but focusing more on “sufficient generalization ability.”

Why does autonomous driving need to skip L3 directly to L4?

The paradigm shift from software rule-based to AI-driven also applies to XPeng’s autonomous driving development.

At the November 2025 Technology Day, XPeng released the second-generation VLA (Vision-Language-Action) intelligent driving model. The upgraded VLA model centers on vision, mimicking human cognition, learning, and observation of the world, eliminating the need for intermediate “language translation,” directly converting visual signals into motion trajectory commands. He Xiaopeng said this reduces information loss, improves reasoning efficiency and response speed, and breaks away from the “video-text” alignment pretraining paradigm, instead training directly on massive real videos.

According to the latest plan, XPeng’s second-generation VLA model will be officially launched in March this year. He Xiaopeng regards this as a “new beginning” for L4-level autonomous driving. He believes that after the second-generation VLA is launched, it will drive sales growth in the third and fourth quarters of this year.

Industry classification divides driving automation into levels 0 to 5: below L3 is assisted driving, L3 is conditional automation, and L4 and L5 are high and full automation, respectively.

Currently, most intelligent connected vehicles are still in L2. In December 2025, the Ministry of Industry and Information Technology announced the first batch of L3 conditional autonomous driving models approved for entry, with two models suitable for urban congestion and highway segments to be tested in designated areas in Beijing and Chongqing.

However, XPeng seeks to leap directly from L2 to L4. This idea is reflected not only in the second-generation VLA model but also in He Xiaopeng’s proposals at the National Two Sessions this year: promoting policies and regulations to bridge L2 to L4, simplifying the L3 intermediate stage, and accelerating technological iteration and large-scale commercialization, which will help China convert its L2 industry advantages into a competitive edge in the L4 autonomous driving era.

He Xiaopeng told reporters that only after legal rules clarify responsibilities at L4 can insurance responsibilities be further determined, and related hardware and software technologies will then evolve toward L4.

Why skip L3? He Xiaopeng explained that software technology has shifted from rule-based to AI-driven, and hardware has been designed with redundancy. Based on this, road traffic regulations should also be forward-looking. If L3 rules are established first, then L4 standards will be delayed. “With current technological growth (speed), plus China’s strength in AI, we should be able to formulate L4 regulations now.”

How can Chinese companies protect their interests abroad when facing bullying?

In the opening letter released on February 24, He Xiaopeng discussed overseas automotive business, expressing the hope to double overseas sales by 2026 and sell 1 million units abroad by 2030, contributing over 70% of profits.

Company data shows that as of December 31, 2025, XPeng Motors has expanded to 60 overseas countries and regions, entering nearly 30 new markets in 2025.

“Compared to many companies, we are still a bit slow in our global expansion. We’ve encountered many difficulties,” He Xiaopeng said in an interview. The internal organizational structure was not originally built specifically for globalization; during the “going abroad” process, they faced issues like incomplete overseas autonomous driving regulations, localization manufacturing, and supply chain challenges. “Now we are enduring pain and joy, learning and understanding many aspects of overseas markets.”

He Xiaopeng noted that many young entrepreneurial companies are “born global,” even prioritizing overseas markets before domestic ones, which is very different from traditional models. In this new context, he believes it is necessary to promote the introduction of laws protecting Chinese companies’ overseas investments. If Chinese private entrepreneurs face unfair treatment or legal issues abroad, Chinese government departments should intervene promptly to coordinate and help resolve, protecting Chinese enterprises’ interests overseas.

Legal experts are also paying attention to this issue. Liu Yanhong, Vice President of China University of Political Science and Law, wrote in a November 2025 article in Legal Daily that some Western countries, under the guise of anti-corruption and free trade, have launched “legal wars,” applying extraterritorial jurisdiction, asset freezes, and financial sanctions against Chinese “going abroad” companies. Accelerating the development of China’s independent legal system for foreign interests protection is necessary to effectively address legal risks in foreign-related law, and to provide practical legal safeguards for Chinese companies’ overseas rights. Liu Yanhong suggested timely drafting key laws such as the Overseas Interests Protection Law.