This week: XPENG and university researchers are spotlighting GroundingPI, a free 4-billion-parameter model whose sole job is answering two questions for robots and cars: which thing, and exactly where is it? You can try it in a browser demo tonight.

Elsewhere: How the model turns a plain-English request into coordinates, what the robot tests showed, the fine print on those scores, and three more things that shipped this week.

The New Prime Signal

Ask a robot to grab a mug and you have quietly asked it two questions. Which object is the mug? And where, precisely, is the handle? Plenty of robot brains handle the first and fumble the second, especially on a cluttered table. On November 4, Chinese EV and robotics company XPENG announced GroundingPI, an open-source visual perception model built by researchers affiliated with XPENG, Peking University, The University of Hong Kong and other institutions. The research paper, model weights, and code are all public, and there is a free online demo if you just want to see what it does.

What happened: GroundingPI is a roughly 4B-parameter model that looks at one image, reads a text instruction, and answers with locations instead of paragraphs: points and bounding boxes, written as coordinates. XPENG says it is trained across 11 categories of perception skills, from object detection and finding "the red car" rather than just "a car" to text recognition (OCR), document layout, app screens, and robotic spatial localization. In the team's own evaluation, it was compared with 44 baseline models across 34 perception benchmarks and posted an average score of 73.68, the highest of the group. A larger general-purpose frontier model in the same comparison scored 71.54. One timing note: the paper appeared September 30 and the weights in October, so this is XPENG's formal introduction, not a fresh upload.

Why it matters: Robots learn new tasks from human demonstrations, which are slow and expensive to collect. The bet is that if the "seeing" is already sharp, practice data can go toward learning how to move. The simulation results point that way. In RoboCasa-GR1 tests, robot learning systems using GroundingPI as their visual backbone reached higher task success rates with 50% of the demonstration data than the comparison models trained with 75%. In RoboTwin 2.0, it beat every mainstream backbone the team evaluated in all four out-of-distribution settings, meaning unfamiliar scenes, by up to 24.8% relative to the strongest one. It was also tried for driving: on the nuScenes dataset it reached an average open-loop trajectory error of 0.296 metres.

The fine print: Every score here comes from the research team. XPENG's own release says the headline comparison "relates to the evaluated perception benchmarks and does not represent a broader assessment of general AI intelligence or reasoning capabilities." The win is not uniform, either. The paper shows the bigger frontier model still well ahead on GUI grounding and some spatial pointing tests, and notes that VisDrone, a drone-footage dataset, "remains challenging." The robot and driving results come from simulation and offline tests; the release says they "do not establish readiness for real-world commercial deployment," and that this is not an announcement of the model going into XPENG's cars or robots. Licensing has a wrinkle: the team's own work is Apache 2.0, but Kimi-derived parts carry the Kimi K3 License, so read the LICENSE file before building a product on it.

We like this one because it is specific. Most AI news is about models that do everything; GroundingPI does one thing, pointing, and argues a robot should get that right before it gets clever. A 4B model edging past a much larger one on its home turf doesn't mean small beats big. It means a specialist can be very good at a narrow job, which, to be fair, is how most of us got hired. And when a company writes its own caveats into the press release, we are inclined to read the rest more generously.

Think of it as a model that answers every question by pointing at the picture.

How it works: Instead of writing a description, GroundingPI writes coordinates. According to the model card, it has 1,000 special location tokens, numbered 0 to 999, that map onto the image. A box is four of them (the top-left and bottom-right corners) and a point is two. Ask for something that isn't there and it answers "None" instead of inventing a spot. Under the hood it pairs a vision encoder from Moonshot's Kimi models with a 4B open language model, with a final reinforcement learning stage that rewards boxes that line up with the right objects. The paper says this format is compact: about 7.6 tokens per box on COCO images, which matters when a robot needs answers fast.

Try it: The easiest route is the GroundingPI demo on Hugging Face Spaces. Upload a photo of your desk or a busy street and describe one thing, the way the model card's example does: "Locate the target referred to by the following description: the red car." The card also lists prompts for reading text and for finding the button to click in a screenshot. Developers can grab the weights on Hugging Face and the code on GitHub, but be realistic: the setup is Linux with Python 3.12 and vLLM, and the tested hardware is data-center GPUs such as NVIDIA H200 and B200, not a laptop. The paper has the full benchmark tables if you want to see where it wins and where it doesn't.