XPENG's Head of Robotics Admits Hardware is 'Bottleneck' as 2026 Production Goal Looms

XPENG's head of robotics, Mi Liangchuan (LC), admits that hardware reliability is currently the primary bottleneck holding back the company’s AI capabilities.
The company is actively building a "data factory" in Guangzhou to solve the scaling challenges of Physical AI, though the specific type of training data remains proprietary.
The production-grade Iron robot will utilize a triple-Turing chip architecture and run VLA 2.0, the same underlying AI stack used in XPENG’s autonomous vehicles.
Reporting reliability issues with first-generation hands, XPENG is pursuing multiple R&D tracks for its next-gen dexterous hardware rather than settling on a single design.
XPENG views the humanoid market as a non-competitive space at this stage, prioritizing the "boundary between human and machine" over immediate rivalry with firms like Tesla or Boston Dynamics.

BEIJING — Humanoids Daily is on the ground in Guangzhou and Beijing this week for XPENG’s 2026 "AI in Motion" media tour. At a technical workshop Mi Liangchuan, the company’s head of AI and robotics, offered a candid assessment of the hurdles facing the Iron humanoid platform. While the company maintains its ambitious 2026 mass production target, Mi suggested that they are currently being held back by the stubborn realities of mechanical engineering.

Mi Liangchuan, head of XPENG Robotics, stands on stage at a technical workshop in Beijing. He wears a black t-shirt and glasses, holding a microphone in his right hand while his left arm is crossed. Next to him on a high round table is a prototype of a white robotic hand and forearm, a bottle of water, and other small electronic components . — Mi Liangchuan (LC) discussing the development of the "Iron" humanoid robot at a 2026 media workshop. The display includes the first-generation dexterous hand prototype, which LC noted is currently being iterated upon to solve hardware reliability issues. Image: Humanoids Daily

Software Speed vs. Hardware Hiccups

Speaking to us directly about the current state of development, Mi acknowledged that XPENG’s AI progress is currently outpacing its hardware. "Surprisingly, AI is not our bottleneck," Mi told the gathered media. "AI is going strong... but all our falling in the past, almost none of them came from AI or the controller. It’s all coming from a little wire somewhere or some signal disconnection".

This "hardware bottleneck" is particularly evident in the development of the robot's extremities. XPENG has previously noted that early factory trials saw hands fail in less than a month. Mi confirmed that the team is still exploring multiple "tracks" for a second-generation hand but has yet to conclude which path will reach the final production line.

The Guangzhou Data Factory

To feed the "Physical AI" models, Mi revealed that XPENG is constructing a dedicated data factory in Guangzhou. While he declined to specify the exact nature of the data being collected, he noted that the "scaling laws" seen in Large Language Models (LLMs) appear to apply to robotics as well.

The strategy relies heavily on "manufacturing synergy" with XPENG’s automotive wing. The latest version of Iron is already running VLA 2.0 (Vision-Language-Action), the same architecture powering the company’s latest autonomous driving systems. By sharing Turing AI chips and navigation models between cars and robots, XPENG hopes to solve the "long-tail" problems of navigation more efficiently than pure-play robotics firms.

A Human-Centric Mission

Beyond the technical specifications—which include an 84-joint architecture and a five-degree-of-freedom waist—Mi framed the project in deeply personal terms. He shared the story of his neighbor, Danny, who was forced into a care facility and recently passed away because his family could not support his needs at home.

"Our motivation to build a humanoid robot to help people is an ultimate mission," Mi said, visibly moved. He envisions the Iron platform eventually serving as a "family member" that allows the elderly or sick to remain in their homes.

Market Outlook: No Room for Rivalry

Interestingly, Mi dismissed the idea of a "humanoid race" against global competitors like Tesla or Unitree. Despite XPENG's recent surging market capitalization, he argued that the potential market is too vast for traditional competition to be the primary focus.

"We don’t treat any team from any country as a competitor because the market is so huge and the industry is still at an early stage," Mi stated. Instead, the company is focusing on "adoption"—the psychological process of humans accepting a machine that looks and moves like them.

As XPENG prepares to break ground on its Guangzhou production base in the first quarter of 2026, the focus has shifted from proving the robot is "real" to ensuring it is reliable enough to start mass producing it.