NederlandsXpeng's small open AI model locates objects better than GPT-6 Astra
Carmaker Xpeng and several universities have released GroundingPI, an open AI model that helps robots and cars see exactly where things are. On 34 tests it scores slightly higher on average than OpenAI's much larger GPT-6 Astra, according to its makers.
The Chinese carmaker Xpeng presented GroundingPI on 4 November. It is an AI model that looks at images and shows exactly where something is. The model was built together with researchers from Peking University, the University of Hong Kong, UC Berkeley and the National University of Singapore, among others. The code and the model weights (the trained "knowledge" of the model) can be downloaded for free.
According to its makers, GroundingPI scores an average of 73.68 per cent on 34 public tests. That is slightly higher than the 71.54 per cent of GPT-6 Astra, the large model that OpenAI released in September. The Xpeng model is much smaller, with about 4 billion parameters (the adjustable numbers of an AI model).
Why a robot must first look carefully
A robot that picks up a cup, or a car that avoids a cyclist, must first know exactly where that cup or cyclist is. According to Xpeng, the large general-purpose AI models were mainly built for the digital world, and they fall short when they have to help a car or robot in the real world. Researchers call the skill of pointing out exactly where something is "grounding".
Xpeng's starting point is that seeing accurately comes first for a smart robot. GroundingPI was therefore trained only to find and point out things, without any examples of robot movements. The model answers with a point or a box on the image. The 34 tests fall into eleven types of task, such as recognising objects, reading text, finding buttons on a screen and pointing out spots for a robot arm. Each type of task counts equally in the average.
The model is built from existing open parts: an image recogniser from the Chinese model Kimi K3 and Alibaba's small language model Qwen3-4B.
Xpeng sees GroundingPI as the "fast eyes" of a robot. A different, larger model is still needed to understand and plan a task. GroundingPI makes sure the robot quickly sees where everything is, so that it can act straight away.
Better with fewer practice examples
Robots often learn a task by watching many examples in which a human guides the robot through the task. Recording such examples is expensive and takes a lot of time.
In a simulation of a kitchen (the RoboCasa-GR1 test), a robot system with GroundingPI as its "eyes" received half of the practice examples. It still succeeded more often in its tasks (28.75 per cent) than all comparable systems that received three quarters of the examples (the best reached 27.75 per cent). According to Xpeng, this is because GroundingPI already masters finding objects, so the robot does not have to learn that itself.
The system also did best in the simulations in situations that the robot had not seen during practice, such as other objects, other bins or a different background. Only in the ordinary kitchen test without such changes did another model, RynnBrain, score slightly higher (39.00 against 37.75 per cent). In a test with images of real car journeys (nuScenes), the predicted route of the car was on average 29.6 centimetres away from the real route, slightly better than the models it was compared with.
Caveats
All results come from the makers themselves and have not yet been measured by others. The robot tests were done in computer simulations, not with real robots. Xpeng itself still calls the work a scientific prototype. The research paper has been online since 30 September and was updated in October. Because the code and model are public, other researchers can now check the results themselves.
Xpeng wants to use the technology for its self-driving cars, robotaxis and humanoid robots. Earlier this year, the company brought its departments for self-driving and smart in-car software together in one AI centre, led by Liu Xianming.
What does this mean for the road to the singularity?
A robot that works in the real world must first be able to see where everything is. GroundingPI shows that a small, open model can keep up with a large model from OpenAI in this, and that a robot with such good "eyes" needs fewer practice examples in simulations. Fewer examples mean cheaper and faster robot training. The next step is to show that this also works with real robots and cars outside the simulation.
What is the technological singularity? Read the explanation →
