XPeng Unveils a Decade-Long “Physical AI” Roadmap: Bridging Bits and Atoms
Want to read in a language you're more familiar with?
XPeng outlines a bold decade-long roadmap for “Physical AI,” spanning L4 autonomous cars, humanoid robots, and consumer flying vehicles—all designed to merge the digital and physical worlds.
On the main stage of GeekPark Innovation Conference 2026, XPeng Chairman and CEO He Xiaopeng delivered one of the event’s most ambitious keynote speeches—laying out a sweeping roadmap for what he calls the coming era of “Physical AI.”
A serial entrepreneur who has lived through the internet, mobile internet, and AI cycles, He argued that the tech world is undergoing a fundamental shift—from the age of digital intelligence to a new paradigm where AI directly shapes the physical world. From XPeng’s end-to-end autonomous driving models to its IRON humanoid robot and next-generation flying vehicles, the company is constructing an ecosystem that tightly couples digital intelligence (bits) with physical systems (atoms).
He predicts that over the next decade, society will move beyond scale effects and network effects, entering what he calls the “Agent Effect”—a new industrial logic shaped by autonomous intelligent entities.
From Network Effects to the Agent Effect
He frames the evolution of industry through the lens of three dominant effects:
- Scale effect in the physical world (bigger production, lower cost)
- Network effect in the internet era (more users, exponential value)
- Agent effect in the Physical AI era (autonomous intelligent systems creating compounding capability)
In this new model, agents—robots, vehicles, digital systems—will learn, cooperate, and execute tasks in increasingly emergent ways. He highlights two emerging phenomena already visible in autonomous driving:
Black Hole Effect: AI continuously absorbs data, compresses knowledge, and produces unexpected capabilities. Example: XPeng’s driving model naturally “creeps” at intersections based on data-driven intuition—not hard-coded rules.
Ant Colony Effect: Decentralized, locally communicating agents demonstrate robust, adaptive intelligence—mirroring how ants coordinate without centralized command.
He argues that this shift will reshape business architectures: while past giants were built on “100,000 employees and 100,000 tools,” and internet companies scaled through “humans + servers,” future enterprises will scale through “humans + tens of millions of AI agents + massive compute.”
Reinventing Autonomous Driving: From Language to Action
He challenges the long-held belief that language-based models are sufficient for world understanding. “Humans don’t learn to swim from reading. We learn by sensing,” he says.
To break through the limitations of L2 assisted driving, XPeng has adopted VLA (Video–Language–Action) models that bypass purely linguistic reasoning, instead grounding AI directly in the physical world. This enables far more natural and capable perception-planning-control loops.
He forecasts two types of autonomous vehicles in the coming five years:
- Robotaxi fleets with fully driverless operation (initially small-scale)
- L4-capable consumer cars requiring a driver in the seat but offering near-full autonomy
XPeng plans to launch three L4-experience vehicles next year, alongside a new Robo-series car and expanded Robotaxi pilots for urban, residential, and indoor parking scenarios.
Why the Future of Embodied AI Is Humanoid: XPeng IRON
XPeng’s flagship embodied-intelligence project is its humanoid robot IRON.
He explains three reasons humanoids are inevitable:
Environment Compatibility: The world is built for humans—narrow hallways, door handles, tools—making humanoid morphology the most adaptable.
Data Availability: Only humanoid robots can directly learn from the vast corpus of human motion data.
Tool Use: Over a million tools are designed for human anatomy; humanoid robots inherit compatibility for nearly all of them.
XPeng has iterated through seven generations of robots (including four quadrupeds) and found non-humanoid forms dramatically less versatile in real homes.
IRON integrates human-like anatomy including articulated shoulders, a multi-joint waist, and a realistic gait system. XPeng’s next milestones include “zero-gravity control posture” and generative motion planning across 70–80 joints—critical prerequisites for mass-market humanoid robots.
The company plans to mass-produce its next-generation robot motion-control system in 2025, backed by XPeng’s full-stack in software, hardware, embedded systems, and manufacturing. R&D spending is expected to reach 11 billion RMB in 2025.
Beyond the Ground: XPeng’s Flying Vehicles
XPeng is also expanding into low-altitude aviation after 12 years of development.
Two vehicle platforms were highlighted:
- “Land Aircraft Carrier” — a modular flying car that stores its folded aircraft inside a vehicle trunk; ideal for tourism flights under 20 minutes, with long-range versions in development.
- A868 eVTOL — a pure aircraft designed for multi-passenger, long-range urban and intercity transport. He believes that in the next decade, “many more people will learn to fly real aircraft.”
The Next Decade of Physical AI
XPeng aims to become a global embodied-intelligence company—building cars, robots, and aircraft as interconnected intelligent agents powered by on-device compute.
Its IRON robot, equipped with three self-developed Turing chips delivering 2,250 TOPS, runs multiple XPeng operating systems (VLT, VLA, VLM) to achieve high-level autonomy without relying on constant cloud connectivity.
In China, deployment will start with commercial scenarios; in Europe and the U.S., industrial applications will take priority. Household robots will follow once costs and reliability mature.
He concludes: “The next decade will bring a wave of Physical AI innovations that transform how we move, work, and live.”
