He Xiaopeng on Physical AI, the IRON Robot and Vehicle Strategy

Zhang Xiaojun interviews XPeng founder and CEO He Xiaopeng about AI, assisted driving, humanoid robots and the company’s vehicle business. His strategy and product plans reflect the interview date, not proven capabilities or investment advice.

The conversation begins with AI use by a CEO. He says teams experiment extensively with AI coding, while he focuses on direction and the overall system rather than only tool details.

He says companies can let teams try tools suited to their work and later integrate what proves useful. A CEO and a product team may use the tools at different depths.

He describes AI coding as a helpful tool for junior programmers at the time and predicts changing skill requirements. For complex systems such as automated driving, he sees infrastructure and system design as essential.

They discuss token consumption. He argues that volume alone cannot capture value; organizations should consider costs alongside employee output, task and business scale.

He mentions computing needs in driving and cockpit teams and distinguishes employees’ model tokens from compute used to train physical-world models. These measures are not interchangeable.

Zhang asks what data would be needed to turn a CEO’s experience into an AI skill. He says every person’s experience has limits that would also appear in such a representation.

He distinguishes AI use in digital businesses from manufacturing. A large organization faces hardware, supply-chain, production and safety constraints that a single usage metric cannot capture.

Looking back over the company’s first decade, he describes developing, manufacturing and selling smart electric vehicles. He believes AI will require the team to reconsider software and hardware design.

On automated-driving research, he contrasts improving existing systems with exploring larger foundation models. Both expanding capabilities and reducing errors are difficult; a best-case demonstration is insufficient.

He criticizes merely attaching AI to old workflows without reconsidering the overall approach. In his view, organizations need to revisit goals, data, processes and products, not just buy tools.

The physical world involves complex continuous data, regulation, supply chains and interactions with people. He argues that extrapolating directly from digital products can underestimate those difficulties.

He predicts the balance of hardware and software value will change and wants organizations to redesign their interaction around AI. This is a strategic judgment, not a verified industry ratio.

He recalls reconsidering the automated-driving roadmap, concerned that local optimization might constrain future capability. The advanced-driving goals discussed are research ambitions, not evidence of a vehicle’s present autonomy level or safety.

Asked for the details of organizational change, he says some work is ongoing and cannot be disclosed. He emphasizes gathering different perspectives while leadership remains responsible for the final bet.

He says key decisions are not necessarily about being more willing to gamble, but about acting in time once a judgment is formed. Different advisers usually see only part of the problem.

On “physical AI,” he emphasizes use cases, regulation, business and people alongside technology. Digital AI expertise does not automatically ensure successful deployment of physical systems.

He expects adoption to differ across company sizes and types. Roles, responsibilities and business models for physical-world tasks still need to be worked out in practice.

Using driving as an example, he says reliably helping people across varied environments is hard. Small teams can innovate, but should not underestimate physical complexity and safety requirements.

On organizational change, he advocates considering processes, teams and products together once the direction is clear. This is his management approach, not a prescription that every company should restructure at the same scale.

He sees combining AI with human health, elder care and whole-company operations as complex problems. A department-level pilot and coordination across an enterprise present different challenges.

Zhang asks whether rapid technological change creates anxiety. He says it can force people to reconsider deeply held assumptions, while any individual sees only part of the problem.

The conversation turns to the development of the IRON robot. He describes the team trying different approaches and encountering both successes and failures along the way.

He says changes around 2023 in his assessment of model capabilities increased the team’s confidence in a robot “brain.” Confidence and investment do not mean a general-purpose robot has been achieved.

He hopes robots can eventually perform practical work and interact appropriately with people. Specific capabilities, reliability and social acceptance still require long-term testing.

He says a quick demonstration is not the same as solving mass production, quality and long-term use. He considers sustained iteration and engineering systems harder challenges.

He explains the choice of a humanoid form by pointing to human environments and interactions. Other tasks may suit other forms, and feasibility must be tested in real settings.

He gives the example of a robot dog potentially bumping furniture in a home and worries that a large hard-bodied robot may feel intimidating. Safety, comfort and size all affect household use.

He mentions the uncanny valley, law and social acceptance. The team wants a form close enough to people for interaction while retaining a distinction; this is a design goal.

Discussing a robot demonstration that drew public skepticism, he says the team disagreed on how to respond. He says they later showed the robot’s structure; this is his account of the episode.

He believes public discussion revealed different expectations for home robots across age groups. He treats those reactions as feedback, not a systematic survey.

Asked about competition and odds, he describes robotics as very difficult and estimates the company’s chances at about 20%. The number is a subjective judgment, not a verifiable probability.

He considers motion control and dexterous manipulation unresolved challenges and compares the field with early automobiles. The analogy illustrates immaturity but cannot predict the pace of progress.

He describes vehicles, robots and related AI as future business tracks, emphasizing hardware–software integration. Specific products and commercialization remain uncertain.

On production challenges, he lists hardware reliability, coordination among models and proof of commercial demand. A demo and reliable manufacturing, service and safety systems are different milestones.

The interview turns to GX and other vehicles. He describes the positioning and plans at the time; specifications, prices and launch details should be checked against official announcements.

He discusses applying ideas from robotics to vehicle interaction, noting that a task requires understanding instructions, surroundings and action steps. These are future concepts, not guarantees about deployed features.

Reflecting on earlier models, he says the organization, product planning and customer understanding have changed. His confidence in a new product is a judgment; actual performance depends on execution and the market.

He says a CEO should focus on strategy and planning, distinguishing sales, profit and long-term value. One metric can reflect a phase without fully determining whether a company is successful.

He cites sales and a technology release to discuss change, but those figures are company observations from the time and do not by themselves prove that one technology caused a sales change.

On supply chains, he compares the complexity of a unified solution with a few partners and individualized solutions with many partners. Different companies may choose different routes, with outcomes still uncertain.

He closes on learning: reading, travel, conversation and hands-on practice should form a feedback loop. He emphasizes testing ideas in practice and revising judgments based on results.