XPeng releases GroundingPI model with superior perception scores against gpt-6 astra

XPeng Group, in collaboration with research teams from Peking University and the University of Hong Kong, has unveiled GroundingPI, a foundational embodied perception model. The research emphasizes a perception-first technical approach, achieving top rankings across 34 public perception benchmarks with approximately 4 billion parameters. When benchmarked against OpenAI's GPT-6 Astra across the same 34 perception metrics, GroundingPI scored 73.68, surpassing Astra's score of 71.54. The model prioritizes precise spatial awareness, utilizing purely perception-based data—without joint action labels—to cover 11 types of capabilities, such as object detection and spatial localization. Initial simulations demonstrate two key advantages for real-world deployment. First, it significantly lowers instruction cost; the system achieves success rates comparable to models requiring 75% instruction data using only 50% with GroundingPI. Second, it exhibits stronger scene generalization, allowing the robot to maintain performance despite variations in material, object type, or background, supporting the concept of cross-station weight reuse.