Xpeng unveiled an embodied perception foundation model named GroundingPI, saying its perception module's overall benchmark score surpassed OpenAI's GPT-6 Astra, according to a November 4, 2026 report. With a parameter scale of about 4 billion, GroundingPI ranked first in a comprehensive evaluation across 34 public perception benchmarks, scoring 73.68 versus 71.54 for GPT-6 Astra, which OpenAI released in September this year, in a comparison weighted equally across 11 capability categories. The model was developed with research teams from universities including Peking University and the University of Hong Kong. In the name, Grounding corresponds to visual localization perception capability and PI stands for Physical Intelligence; the code and model weights have been open-sourced. Xpeng said existing general-purpose large models are designed for digital scenarios and face an adaptation gap in physical scenarios such as autonomous driving and robotics. GroundingPI does not use robot data with joint action labels during training, relying entirely on pure perception data without action labels across 11 capability categories including general detection, text recognition, and robot spatial localization. It aims to provide a perception-native foundation for the fast system rather than replace the slow system's general large model capabilities. The research is at the scientific prototype stage, and a system using GroundingPI as its visual foundation can exceed the execution success rate of comparison models using 75% of the demonstration data while using only 50% of the data.