Xpeng Beats OpenAI on Perception Benchmarks with 4-Billion-Parameter GroundingPI
In this article (8)
Electric vehicle maker Xpeng has launched an embodied perception foundation model named GroundingPI, reporting an overall benchmark score that surpassed OpenAI’s frontier model GPT-6 Astra across 34 public perception tests.
Built on a parameter scale of about 4 billion, the perception-first architecture ranked first in comprehensive evaluations and demonstrated higher data processing efficiency in robot simulation tasks. Xpeng developed the model in partnership with research teams from Peking University and the University of Hong Kong, aiming to significantly reduce the cost of manually collecting robot demonstration data.
In a direct comparison across 11 capability categories, GroundingPI recorded a score of 73.68, edging out the 71.54 achieved by GPT-6 Astra, which OpenAI released in September this year. The code and model weights have been open-sourced.
Benchmark Scores and Technical Design
Existing general-purpose large models are primarily designed for digital scenarios and face an adaptation gap when applied directly to physical scenarios such as autonomous driving and robotics, Xpeng said. GroundingPI proposes that the primary prerequisite for embodied intelligence is accurately perceiving spatial positions.
Therefore, the model does not use robot data with joint action labels during the training phase, relying entirely on pure perception data without action labels to cover 11 capability categories including general detection, text recognition, and robot spatial localization.
Embodied intelligence is typically divided into a fast system (System 1) responsible for high-speed perception and action execution, and a slow system (System 2) responsible for task understanding and planning. GroundingPI aims to provide a perception-native foundation for the fast system, rather than replacing the general large model capabilities of the slow system, Xpeng said.
“The research is currently at the scientific prototype stage, but its simulation results have provided validation for 2 core challenges in scaled deployment.”
Halving Data Collection in Warehouse Trials
The research is currently at the scientific prototype stage, but its simulation results have provided validation for 2 core challenges in scaled deployment. The first is significantly reducing the cost of manually collecting robot demonstration data.
A system using GroundingPI as its visual foundation can exceed the execution success rate of comparison models using 75% of the demonstration data while using only 50% of the data, because it directly takes on the task of locating objects in traditional Vision-Language-Action (VLA) models.
The second is bringing stronger scenario generalization. In high-density simulation scenarios such as bulk material sorting and bin counting, the system maintains stable performance when dealing with changes in materials, object categories, and backgrounds, providing support for reuse across different workstations.
Centralising Machine Intelligence
GroundingPI represents an important advancement for the Chinese EV maker in building its physical AI system. Xpeng has continuously increased its bets on physical AI in recent years, hoping to drive its smart driving systems, humanoid robots, and robotaxi businesses.
Earlier this year, Xpeng merged its autonomous driving center and smart cockpit center into a general intelligence center led by AI expert Liu Xianming , pooling resources to provide unified AI middle-platform foundation support for its automotive and robotics businesses.
An Xpeng Iron humanoid robot was previously put on display at the April 2026 Beijing Auto Show.
Questions & Answers
Q.How did GroundingPI perform against OpenAI's frontier model?
How did GroundingPI perform against OpenAI's frontier model?
In a direct evaluation across 11 capability categories and 34 public tests, GroundingPI recorded a score of 73.68, edging out the 71.54 achieved by OpenAI's GPT-6 Astra. The four-billion-parameter architecture also demonstrated higher data processing efficiency during robot simulation tasks.
Q.How does the model reduce the need for robot demonstration data?
How does the model reduce the need for robot demonstration data?
The model does not use robot data with joint action labels during training, relying entirely on pure perception inputs. Consequently, robot systems using GroundingPI can surpass the execution success rates of traditional models whilst cutting required demonstration data to 50 per cent.
Q.What role does GroundingPI play in Xpeng's broader technology strategy?
What role does GroundingPI play in Xpeng's broader technology strategy?
GroundingPI is designed as a perception-native foundation for high-speed systems, supporting Xpeng's smart driving, robotaxis, and humanoid robots. It fits into a wider corporate push that merged the carmaker's autonomous driving and smart cockpit units into a single general intelligence centre.
Q.Is GroundingPI ready for commercial use?
Is GroundingPI ready for commercial use?
The model is currently at the scientific prototype stage, though Xpeng has open-sourced both its code and weights. It was created in partnership with Peking University and the University of Hong Kong to address core challenges in scaling robot deployments.
Reader pulse
Can Xpeng translate its perception AI lead into commercial hardware?
22,196 votes so far