XPENG-linked researchers release free AI model that helps robots point to objects they see
The team says its 4-billion-parameter GroundingPI beat larger models on its own tests.
Researchers tied to Chinese electric-car maker XPENG released a free AI model Wednesday meant to help robots and cars find exactly where things are in what they see.
The model, GroundingPI, looks at an image and answers with coordinates: a point or a box marking a spot, such as "the handle of the cup on the right." That is called grounding. A chatbot can describe a kitchen, but a robot arm needs to know where to reach.
The team includes researchers from XPENG, Peking University and the University of Hong Kong. The model has about 4 billion parameters, the adjustable numbers inside an AI model, which is small by current standards. The code and model files can be downloaded from GitHub and Hugging Face.
XPENG said the model scored 73.68 across 34 tests, ahead of 71.54 for OpenAI's GPT-6 Astra, in the team's own evaluation. In simulated robot tasks, it said robots using it did better while trained on half as many human demonstrations. Those tests were run in software simulations, not on real robots in homes or factories.
XPENG said the research "remains exploratory" and the results "do not establish readiness for real-world commercial deployment." The research paper has not been peer reviewed.
Source: XPENG press release, Nov. 4, 2026; GroundingPI paper (arXiv 2609.39601)