RoboBoom
Humanoids Summit· March 10, 2026· 18:57

XPENG IRON: The Most Human-Like Catwalk Humanoid with Jian Yao

Jian Yao, an AI engineer at XPENG, describes the IRON humanoid’s design goals and AI architecture at Humanoids Summit.

In this talk

  1. 1:23

    Building a human-like robot

    Yao says XPENG aims for IRON to resemble people in appearance, movement, independent navigation, communication, and actions. He stresses that assembling the hardware also requires solving thermal management, power, operating-system, and sensor-calibration challenges.

  2. 4:46

    Coordinating the robot’s skills

    Yao describes VLT as an agent-like brain that interprets commands and coordinates navigation and manipulation. For example, it can send IRON to fetch a water bottle, respond if the bottle falls, and then resume the trip. VA 2.0 processes inputs such as video, text, touch, and the robot’s state to produce outputs for movement and manipulation.

  3. 9:09

    Designing for on-device limits

    XPENG collects manipulation, navigation, and communication examples through a data factory. Yao says battery, heat, and compute limits make it hard to run large models while still producing actions at high frequency. XPENG therefore co-designs its model architecture and inference engine for its hardware and trains multimodal models from scratch.

  4. 13:32

    Product plans and open questions

    Yao gives the end of 2026 as XPENG’s target for commercialization and mass production. He says dealerships are one intended use case and that product-level reliability is expected. XPENG may consider releasing some foundation models, but its hardware-specific design could limit compatibility with other chips; Yao also describes real-time control as necessary to close small gaps in whole-body actions.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks