# XPeng Group and University Partners Release GroundingPI Pretrained Model, Topping GPT-6 Astra on Perception Benchmarks

XPeng Group, together with research teams from Peking University, the University of Hong Kong and other universities, released GroundingPI, an embodied perception foundation model with about 4 billion parameters that ranked first on a combined evaluation of 34 public perception benchmarks\.

By China EV Signal
Published: 2026-11-04T13:58:38.813Z
Updated: 2026-11-04T13:58:38.815Z
Canonical article: https://chinaevsignal.com/news/xpeng-groundingpi-pretrained-model-perception-benchmarks

## Key takeaways

- GroundingPI, with about 4 billion parameters, ranked first in a combined evaluation of 34 public perception benchmarks\.
- On the same 34 benchmarks, averaged equally across 11 perception capability categories, GroundingPI scored 73\.68 and GPT-6 Astra scored 71\.54\.
- GroundingPI was trained entirely on perception data without action labels, covering 11 perception capability categories, and its code and model weights have been fully open-sourced\.

## GroundingPI tops perception benchmark evaluation

XPeng Group, together with research teams from Peking University, the University of Hong Kong and other universities, released GroundingPI, an embodied perception foundation model\. With about 4 billion parameters, GroundingPI ranked first in a combined evaluation of 34 public perception benchmarks\.

The research team compared GroundingPI with OpenAI's frontier model GPT-6 Astra on the same 34 perception benchmarks\. Averaged equally across 11 capability categories, GroundingPI scored 73\.68 and GPT-6 Astra scored 71\.54\.

## Training uses perception data without action labels

GroundingPI does not use robot data with joint action labels during training\. It is trained entirely on pure perception data without action labels, covering 11 perception capability categories including general detection, dense small objects, referring localization, text recognition, document layout, interface elements and robot spatial localization\.

In a system using GroundingPI as its visual base, 50% of demonstration data was enough to exceed the success rate of a comparison model using 75% of the data\.

## Code and model weights open-sourced

The related code and model weights have been fully open-sourced\. XPeng Group said it welcomes industry discussion and verification\.

The research is currently at the scientific prototype stage\. XPeng Group noted that models on the industry's two technical routes are both iterating quickly, with an update cycle of about half a year\.

## Company coverage

- [XPeng news](https://chinaevsignal.com/manufacturers/xpeng)

## Original publisher

小鹏汽车. Published: 2026-11-04T09:11:14.000Z