On Nov. 4, 2026, XPENG published a Guangzhou-dated press release under the heading “XPENG Researchers and Academic Partners Introduce GroundingPI, an Open-Source Visual Perception Model for Physical AI.” The note highlights open-source perception research XPENG has supported, not a commercial product launch. The paper, weights and code were already public; this release spotlights results the team posted earlier. For related coverage, see OpenAI Teams With Mazda to Test AI on Vehicle Design and Tencent Releases Hy4 With Vision and Computer Use.
Key facts
| Item | Detail |
|---|---|
| Announced | Nov. 4, 2026 |
| By | Researchers affiliated with XPENG, Peking University, The University of Hong Kong and others |
| Size | About 4B parameters |
| Benchmarks | 73.68 vs GPT-6 Astra 71.54 across 34 benchmarks, team’s evaluation |
| Availability | Paper, weights and code public |
| Status | Exploratory research, no product commitment |
What was announced
GroundingPI is described as an open-source visual perception foundation model created by researchers affiliated with XPENG, Peking University, The University of Hong Kong and other academic institutions. XPENG has supported the development. The release states plainly that “The project represents an academic research contribution supported by XPENG and its research collaborators. The results do not constitute an announcement of integration into XPENG’s commercial products or a commitment to future deployment.”
The model carries approximately 4 billion parameters. Per the GitHub README, the paper went on arXiv 2026-09-30 and weights landed on Hugging Face 2026-10-01, so Nov. 4 is an announcement of research already published.
What GroundingPI does
GroundingPI targets physical-AI perception: locating objects and regions in images with spatial coordinates, including points and bounding boxes. Capability categories named in the release include object detection, dense object localization, referring expression grounding, OCR, document layout understanding, user interface element recognition and robotic spatial localization. An example in the release is identifying “the handle of the cup on the right.”
The perceptual model’s training does not rely on robot action labels; the release positions GroundingPI as a visual foundation for downstream action-learning systems. The GitHub README describes the stack as a MoonViT-V2 (Kimi K3) visual encoder, a multimodal projector and a Qwen3-4B language backbone.
The benchmark result and its limits
In the research team’s evaluation, GroundingPI was compared with 44 baseline models across 34 public perception benchmarks spanning 11 capability categories, with equal weighting across categories. GroundingPI scored 73.68 versus 71.54 for GPT-6 Astra. XPENG’s release adds: “The comparison relates to the evaluated perception benchmarks and does not represent a broader assessment of general AI intelligence or reasoning capabilities.”
Robotics and driving tests
Downstream tests cited in the release stay in simulation or offline settings. In RoboCasa-GR1 simulation, systems using GroundingPI as a visual backbone achieved higher task success with 50% of the demonstration data than the evaluated comparison models trained with 75%. In RoboTwin 2.0 simulation, GroundingPI was stronger than the evaluated mainstream visual backbones across four out-of-distribution settings.
On nuScenes, used as a visual backbone, GroundingPI reached an average open-loop trajectory prediction error of 0.296 metres in the reported evaluation. XPENG adds that “The research remains exploratory. Its downstream results are based on simulation and offline evaluations and do not establish readiness for real-world commercial deployment.”
How to get it and licensing
The paper, model weights and supporting code are publicly available, subject to the applicable open-source and third-party licensing terms, at groundingpi.github.io, arXiv 2609.39601, GitHub and Hugging Face. Original contributions use Apache License 2.0; Kimi-derived material follows the Kimi K3 License noted in the README and model card.
What it means
Analysis (ours). The release positions GroundingPI as a perception-first open foundation for physical AI rather than a general assistant. The GPT-6 Astra gap is meaningful only inside the team’s 34-benchmark bundle; it does not settle broader capability debates. Simulation results hint at better data efficiency when GroundingPI backs robot policies, but XPENG’s product disclaimer leaves fleet or factory deployment off the table for now. Openly released weights (Apache 2.0 for original contributions, Kimi K3 License for Kimi-derived parts) plus a documented MoonViT-V2 and Qwen3-4B stack give researchers a reproducible checkpoint independent of XPENG’s Nov. 4 publicity push.
What to watch
Independent replication of the 34-benchmark score, real-robot evaluations beyond simulation, and any follow-on integration studies from XPENG or partners despite the stated lack of product commitment.