XPeng GroundingPI: open-source 4B robot perception model it says beats GPT-6 Astra on benchmarks
XPeng unveiled GroundingPI, built with university partners including Peking University, the University of Hong Kong and UC Berkeley. It's a roughly 4-billion-parameter model that finds exactly where objects are in an image, the kind of perception robots and cars need before they can act. The code and weights are open-source and first went online in early October.
XPeng says it ranked first across 34 public perception benchmarks, averaging 73.68 against 71.54 for OpenAI's larger GPT-6 Astra. It was trained only on perception data, with no robot action labels. Used as the vision backbone in a simulated robot manipulation benchmark, it beat rival backbones trained on 75% of the demonstrations while using only 50%.
Why it matters
XPeng calls it a scientific prototype. The measured robot results come from simulation benchmarks, not real-world robot tests, and the benchmark scores are the team's own.
About the footage: Official GroundingPI demo video: real robot arms at a tabletop with the model's detection boxes overlaid. It doesn't say whether GroundingPI was controlling the arms; the measured robot results are from simulation benchmarks.
The source post
Post by XPENG on X.
View the post on X
Sources
- XPENG on X
- CnEVPost: Xpeng unveils GroundingPI pre-trained model
- GroundingPI project page
- GroundingPI on Hugging Face
New to robots? Read our guide: How to tell if a robot video is real: CGI, sped up, or teleoperated?.
Summarized by Servospace from the sources above, drafted with AI help and checked against those sources before publishing. Claims are the companies' or outlets' unless we say otherwise. How we work.