Gasgoo Munich- XPENG, in partnership with teams from Peking University and the University of Hong Kong, has released GroundingPI, a foundation model for embodied perception. With roughly 4 billion parameters, the model topped a comprehensive evaluation of 34 public perception benchmarks. It scored 73.68 points, beating OpenAI's frontier model, GPT-6 Astra, which scored 71.54.
Image source: XPENG
Physical AI and embodied intelligence are driving the current evolution of the AI industry. However, general large models are primarily geared toward digital scenarios. They face an adaptation gap when applied directly to physical environments. To bridge this, GroundingPI adopts a "perception-first" approach. During training, it avoids using robot data labeled with joint actions. Instead, it relies on pure perception data covering 11 capabilities. These include general detection, dense small-target recognition, and referential localization. This reinforces the model's visual grounding. It sharpens its ability to identify not just what an object is, but where it is in the frame. The model is positioned as a perception-native base for the "fast system". It functions as System 1 of embodiment, rather than replacing the general capabilities of the "slow system". (System 2).
Simulations show that a system built on GroundingPI achieves a higher success rate with just 50% of demonstration data than comparison models using 75%. This suggests a significant reduction in manual teaching costs, alongside stronger generalization across varying materials and backgrounds. The research remains in the prototype stage, though the team has open-sourced the relevant code and model weights.
