XPENG IRON humanoid robot receives a capability upgrade. Powered by the second-generation VLA large model, it can predict human movements up to six seconds in advance, significantly enhancing the flexibility of human-robot interaction.
A common issue with many humanoid robot interactions is that they require humans to follow a fixed rhythm. If a person changes their movement unexpectedly, the robot often fails to respond adequately.
Unlike previous models that would execute a fixed set of movements after recognizing gestures, XPENG IRON continuously observes the person in front of it, re-inferring from the visual input every 0.2 seconds.
In demonstration scenarios, IRON can easily perform simple actions such as handshakes, fist bumps, and high-fives. Even if a person suddenly changes their move mid-action—such as clenching their fist or shifting the high-five position—the robot can instantly adjust its posture and switch hands to cooperate, rather than freezing in place.
This predictive capability stems from XPENG's physical AI foundation. This technology shares its origins with the predictive capabilities used in XPENG's autonomous driving systems and is deployed directly on the robot itself, eliminating the need for cloud-based remote control.
By transferring the ability to understand spatiotemporal changes from automobiles to humanoid robots, the robot can not only interpret the current scene but also deduce what will happen in the immediate future.
In terms of hardware, IRON is equipped with three self-developed Turing AI chips, delivering a total computing power of 2250 TOPS. This is sufficient to support high-frequency real-time inference, ensuring no noticeable delay in motion response.
For humanoid robots to truly integrate into daily life, they must go beyond performing fixed scripts and learn to understand human behavior and anticipate human intentions.
This upgrade marks a significant step forward; the robot no longer just mechanically executes commands but has started learning to understand human behavioral logic.
Last month, while discussing the robotics business, He Xiaopeng revealed that developing robots is approximately 20 times more difficult than manufacturing cars, stating, "It is truly very challenging."
He Xiaopeng stated, "Basically, our robot's hardware, software, design, and supply chain are all in-house. I've been reluctant to showcase our robots in the past mainly because they couldn't be mass-produced. Many robot capability demonstrations on the market are edited before being released."