XPeng has upgraded its second-generation VLA model, which now powers the IRON humanoid robot. The robot can now understand temporal sequences and predict human actions within a six-second window. The system performs a new inference cycle every 0.2 seconds to maintain a real-time feedback loop between visual input and physical movement.
The IRON robot features a human-like design with a spinal structure, bionic muscles, flexible skin, and a 3D curved display for a head. Its hands offer 22 degrees of freedom. The hardware is equipped with three Turing AI chips and utilizes a combination of VLT, VLA, and VLM models to manage dialogue, navigation, and physical interaction.
The company plans to begin mass production of the IRON robot by the end of 2026. Deliveries are scheduled to commence in the first half of 2027. Initial deployment will focus on commercial environments such as retail stores and shopping malls, with future expansion planned for industrial and domestic applications.
