Xpeng releases GroundingPI perception model outperforming GPT-6 Astra

Xpeng has unveiled GroundingPI, an embodied perception foundation model with a parameter scale of about 4 billion that outperformed OpenAI's GPT-6 Astra across 34 public perception benchmarks. Developed in collaboration with Peking University and the University of Hong Kong, the model focuses on visual localization and spatial perception without relying on joint action labels. Xpeng has open-sourced the code and model weights as it works to integrate the perception-native foundation into its autonomous driving, humanoid robot, and robotaxi businesses. Open article →