By running its car-derived prediction model on IRON's own chips, XPENG is turning driving-AI spending into robot capability ahead of deliveries planned from the first half of 2027.
XPENG says the second-generation VLA model behind its driver-assistance system now runs on its IRON humanoid robot, letting the machine anticipate a person’s movements up to six seconds ahead. The company announced the upgrade on October 22, according to MyDrivers and IT Home. VLA stands for vision-language-action: a model that takes in what a camera sees and turns it into physical actions.
In the demonstrations described by both outlets, IRON shakes hands, bumps fists and exchanges high fives. When the person changes the gesture partway through, by suddenly making a fist or moving the hand for a high five, the robot adjusts rather than finishing the original move. IRON recalculates from its camera view about every 0.2 seconds, so the person does not have to follow a fixed rhythm.
What is new is where the model runs and where it came from. MyDrivers reported that the model is deployed on the robot itself rather than relying on remote control from the cloud, and that the prediction ability shares its origin with XPENG’s driving prediction. IT Home described both as resting on the same physical-AI foundation. No independent test of the robot’s behavior has been published.
A prediction feature from XPENG’s cars moves into its robot
The six-second horizon first appeared in XPENG’s cars. In late August, XPENG’s second-generation VLA received its first major upgrade, which added a predictive world model that anticipates what will happen over the next six seconds, Science and Technology Daily reported. The company’s announcement, as quoted by IT Home, made the link explicit: last month the model learned to understand time, and this month it arrived on IRON to understand a person’s next step.
XPENG has now moved the same predictive capability from its cars into a humanoid, about two months after the car upgrade. The hardware is also shared. IRON carries three of XPENG’s in-house Turing AI chips with 2,250 TOPS of combined computing power, according to MyDrivers, the same total TechNode reported when the first IRON came off XPENG’s production line on September 8.
The upgrade comes with a production schedule attached
IT Home’s report restated XPENG’s plan: mass production of the humanoid by the end of 2026, deliveries from the first half of 2027, starting in stores and shopping malls before moving to industrial and home settings. TechNode reported on October 13 that XPENG plans to begin deliveries in 2027 and roll the robot out in Europe later that year, a route EastFrontier covered on October 14.
Chairman He Xiaopeng has framed the robot as a much harder product than a car. MyDrivers re-quoted comments he made last month that building robots is roughly 20 times as difficult as building cars, and that XPENG controls IRON’s hardware, software, design and supply chain. He also said he had long been reluctant to show the robot because it could not yet be mass-produced, and that many capability demonstrations on the market are edited videos. In August, The Paper reported him saying the gap with cars was at least 20 times, and that monthly robot output could rise quickly to several thousand units next year depending on demand.
Those remarks set the standard XPENG has invited for its own robot. A clip of a handshake that adapts mid-motion is a narrower claim than reliable behavior across thousands of interactions in a busy mall, and the October 22 material gives no failure rates or test conditions.
Store deployments will test the claim at volume
The production version of IRON has 76 degrees of freedom, including 21 in each hand, TechNode reported in September, when XPENG said more than 80% of the core processes on its robot line were automated. The company has not said how many units it will build by year-end.
The first robots in XPENG’s own stores, expected in the first half of 2027, will be the earliest public test of whether the six-second prediction holds up with customers who do not follow a script. Before then, the next marker is whether XPENG confirms year-end output figures, or publishes the robot’s performance outside staged demonstrations.