XPeng Upgrades IRON Humanoid Robot With Real-Time VLA Inference

XPeng has upgraded its IRON humanoid robot with a second-generation Vision-Language-Action (VLA) model, enabling edge inference every 0.2 seconds and action prediction up to 6 seconds ahead. This upgrade allows the robot to seamlessly handle dynamic human interactions like handshakes and high-fives, adjusting instantly to unexpected movements. This breakthrough addresses one of the most persistent bottlenecks in humanoid robotics: real-time adaptation to unstructured and unpredictable human behavior. By moving advanced temporal understanding and future action prediction to the edge, it bridges the gap between robotic execution and human-speed social dynamics. The robot re-evaluates visual scenes and performs reasoning every 0.2 seconds while continuously predicting human motion trajectories 6 seconds into the future. XPeng plans to mass-produce IRON by the end of 2026, targeting commercial storefronts before expanding into industrial and home environments.