XPENG and academic partners release open-source GroundingPI embodied perception model

XPENG, with Peking University and The University of Hong Kong, releases GroundingPI, a roughly 4B-parameter pretrained model for embodied perception.

The model is trained only on perception data without action labels. XPENG says it ranks first across 34 public perception benchmarks, scoring 73.68 overall against 71.54 for OpenAI's GPT-6 Astra under equal weighting of 11 capability categories.

The source code and model weights are open-sourced. XPENG says the research is at the research-prototype stage, with an experimental design that points toward real-world deployment in embodied intelligence.