Xpeng Beats OpenAI on Perception Benchmarks with 4-Billion-Parameter GroundingPI

Electric vehicle maker Xpeng has introduced a foundation model called GroundingPI. It is designed to help robots understand what they see and where things are. Xpeng says the model scored higher than OpenAI's advanced model, GPT-6 Astra, in 34 public tests that measure how well AI systems understand their surroundings.

The model has about 4 billion parameters. It ranked first in overall tests and processed data more efficiently in robot simulations. Xpeng developed it with research teams from Peking University and the University of Hong Kong. Their goal is to greatly reduce the cost of collecting robot training examples by hand.

In a direct comparison covering 11 skill areas, GroundingPI scored 73.68. OpenAI's GPT-6 Astra scored 71.54. OpenAI released GPT-6 Astra in September this year. Xpeng has made the model's code and weights freely available.

Test Scores and How the Model Works

Xpeng said that general-purpose AI models are mainly designed for digital tasks. This can make them less suitable for use in the physical world, such as in self-driving cars and robots. GroundingPI is based on the idea that robots must first understand exactly where things are in space.

For this reason, the model is trained without robot data that includes labels describing joint movements and actions. Instead, it uses only data about what things look like and where they are. It learns 11 skill areas, including finding objects, reading text, and identifying the positions of objects for robots.

Embodied intelligence is usually divided into two systems. The fast system, called System 1, handles quick perception and actions. The slow system, called System 2, understands tasks and makes plans. Xpeng says GroundingPI is designed to give the fast system a strong foundation for understanding its surroundings. It is not meant to replace the general AI abilities of the slow system.

Using Less Data in Warehouse Tests

The research is still at the scientific prototype stage. However, its simulation results provide evidence that it may help solve two important problems when robots are used on a larger scale. The first is reducing the cost of collecting robot training examples by hand.

A system that uses GroundingPI to understand visual information can achieve a higher task success rate than comparison models while using 75% of their demonstration data, with the reported data requirement also described as only 50% of the data. This is because GroundingPI directly handles the task of finding and locating objects, which is usually part of traditional Vision-Language-Action (VLA) models.

The second benefit is that the system can work well in more types of situations. In busy simulation tasks, such as sorting large amounts of material and counting objects in storage bins, it continues to perform steadily even when the materials, object types, or backgrounds change. This could make it easier to use the same system at different workstations.

Bringing AI Development Together

GroundingPI is an important step for the Chinese electric vehicle maker as it builds its physical AI system. In recent years, Xpeng has invested more in physical AI. It hopes to use this technology in its smart driving systems, humanoid robots, and robotaxi services.

Earlier this year, Xpeng brought its autonomous driving center and smart cockpit center together into one general intelligence center. AI expert Liu Xianming leads the new center. The aim is to combine resources and build a shared AI foundation that supports both its car and robotics businesses.

An Xpeng Iron humanoid robot was also displayed at the Beijing Auto Show in April 2026.