Xiaopeng and Partners Release GroundingPI Embodied Perception Foundation Model

Xiaopeng Group, alongside Peking University and Hong Kong University, jointly released the GroundingPI embodied perception foundation model with about 4 billion parameters. It utilizes a perception-first training approach and ranks first across 34 public perception benchmarks, narrowly outperforming GPT-6 Astra. This model addresses a critical bottleneck in physical AI by focusing heavily on spatial grounding without requiring joint-action labels during training. By significantly lowering the data demonstration costs and improving environmental generalization, it paves the way for more efficient real-world robot deployment. GroundingPI features around 4 billion parameters and achieves an average score of 73.68 across 11 categories of perception benchmarks, compared to 71.54 for GPT-6 Astra. It relies entirely on action-free perception data covering tasks such as generic detection, small object localization, and text recognition.