XPeng announced on October 22 that its second-generation VLA model has been upgraded and deployed on-device on its humanoid robot IRON. After the upgrade, IRON re-runs its reasoning every 0.2 seconds based on what it currently sees, while also predicting how its actions will unfold over the following 6 seconds.
XPeng's own framing of the upgrade: last month, the second-generation VLA began to understand "time"; this month, applied to IRON, it begins to understand "your next move".
Keeping up when the other person changes their move
The official demos focus on close-range interaction. IRON can shake hands, bump fists and high-five with people, and it adjusts when a person improvises: if an outstretched palm suddenly closes into a fist, it switches to a fist bump; if the high-five target shifts, it uses the other hand to meet it.
According to XPeng, IRON keeps observing, judging in real time and adjusting continuously, without waiting for a full motion to finish before reacting. ITHome's report summed up the difficulty of these scenarios bluntly: "people don't make their moves at a robot's pace."
The older approach has the robot recognize a gesture first, then call up a whole pre-set motion to execute, and during execution it is largely blind to changes around it. With a higher re-reasoning rate, the motion is cut into very short segments, and each one can be corrected against the latest image.
The two numbers: 0.2 seconds and 6 seconds
Taken together, the two figures give a rough idea of the system's cadence. Re-reasoning every 0.2 seconds means five times per second; across a 6-second prediction window, the model is invoked about 30 times, each time updating its view of the next few seconds and then taking only the nearest short slice to execute.
A prediction window much longer than the execution segment has an upside: motion does not stutter, and the robot knows it is in the middle of a longer action. The downside is that every inference must finish within 0.2 seconds, so model size and on-device compute are both constrained by that deadline. XPeng has not disclosed the parameter count of the second-generation VLA deployed on IRON, nor any measured distribution of inference latency.
For comparison, human simple reaction time to a visual stimulus is typically around 0.2 seconds. IRON's re-reasoning interval is in the same range, which looks adequate at first glance for relatively slow interactions such as handshakes and high fives. Whether a 0.2-second cadence can keep up with faster actions, such as swatting a flying object or catching a thrown item, did not come up in this demo.
On-device deployment also has trade-offs: inference runs on the robot itself with no cloud round trip, so network jitter does not directly slow its reactions, at the cost of fitting the model onto the chips inside the body.
The hardware under IRON
IRON was unveiled by XPeng at its Tech Day on November 5, 2025. The body uses a human-like spine, bionic muscles and fully covered flexible skin; the head is a 3D curved display, the shoulders have a bionic structure, and the dexterous hand has 22 degrees of freedom.
For compute, it carries three XPeng-designed Turing AI chips, and on the software side a combination of VLT, VLA and VLM supports conversation, walking and interaction. The second-generation VLA is built on the same physical AI foundation model at XPeng; by the company's account, understanding of "time" was added last month first, and only this month did it land on IRON.
Mass-production timeline
XPeng plans to reach mass production of humanoid robots by the end of 2026 and begin deliveries in the first half of 2027, starting with commercial settings such as showrooms and shopping malls, then expanding to industrial and home use.
In the same week, Unitree released its humanoid H2-X, which it says offers up to about 8 hours of battery life, so the pace of humanoid launches in China remains brisk. Each company's demos look more and more like everyday interaction, but whether the same reaction speed holds up among real crowds after mass production is something XPeng has published no field-test data on. That will only become visible once the first robots are in stores.
Sources: XPeng official announcement, ITHome, CocoLoop; XPeng's announcement was checked for the re-reasoning interval, prediction window and mass-production and delivery plans.