GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives

Hugging Face Daily Papers Papers
grounding foundation-model physical-intelligence vision-language-action robotics autonomous-driving

Summary

GroundingPI：一个 4B 参数的视觉定位基础模型，以量化坐标生成点与框，在 34 个定位基准上以 73.68% 平均成绩超越 GPT-6 Astra 等 44 个基线，并在机器人操作与自动驾驶下游任务中显著提升性能。

Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control. Yet vision-language-action (VLA) and world-action models (WAMs) take perception from general-purpose vision-language and video-generation backbones, which still fail in these settings. We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI establishes a new state of the art, averaging 73.68%, above the larger GPT-6 Astra (71.54%). As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. On RoboTwin 2.0, it outperforms every mainstream backbone we evaluate in all four out-of-distribution settings, by up to 24.8% relative to the strongest backbone. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m. We systematically analyze GroundingPI's pretraining in scale and data composition. Downstream autonomous driving and robotic manipulation improve as the pretraining is scaled. Analyzing the data recipe across these 11 perceptual capabilities shows dense grounding's substantial benefits for both, and OCR's potential as a catalyst for perceptual learning. These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.
Original Article
View Cached Full Text

Cached at: 10/13/26, 06:44 AM

Paper page - GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives

Source: https://huggingface.co/papers/2609.39601 Published on Sep 30

#2 Paper of the day Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

Precisegroundingmatters.Itspecifieswhichobjectisthetargetandwherethatobjectis,eveninclutterandfortinyobjects,andithastobefastenoughforclosed-loopcontrol.Yetvision-language-action(VLA)andworld-actionmodels(WAMs)takeperceptionfromgeneral-purposevision-languageandvideo-generationbackbones,whichstillfailinthesesettings.WeintroduceGroundingPI,a4Bgroundingfoundationmodelthatgeneratespointsandboxesasquantizedcoordinatesinasharedvocabulary.Trainingcombinesmultimodalandspatialpretraining,supervisedfine-tuning,andreinforcementlearningwithGRPO,usingsupervisionfrompublicdatasetsanddedicateddataengines.Against44baselinesacross34groundingbenchmarksspanning11perceptualcapabilities,GroundingPIestablishesanewstateoftheart,averaging73.68%,abovethelargerGPT-6Astra(71.54%).Asadownstreamvisualbackbone,GroundingPIimprovesperformanceonroboticmanipulationandautonomousdriving.OnRoboTwin2.0,itoutperformseverymainstreambackboneweevaluateinallfourout-of-distributionsettings,byupto24.8%relativetothestrongestbackbone.OnRoboCasa-GR1,GroundingPItrainedwith50%ofthedemonstrationsoutperformsthosebaselinestrainedwith75%.OnnuScenes,usedasthevisualbackbone,GroundingPIattainsanaverageopen-loopL2errorof0.296m.WesystematicallyanalyzeGroundingPI’spretraininginscaleanddatacomposition.Downstreamautonomousdrivingandroboticmanipulationimproveasthepretrainingisscaled.Analyzingthedatarecipeacrossthese11perceptualcapabilitiesshowsdensegrounding’ssubstantialbenefitsforboth,andOCR’spotentialasacatalystforperceptuallearning.Theseresultssupportgroundingasaperceptualfoundation,anddedicatedperceptualpretrainingasapromisingdirectionforfoundationmodelsofphysicalintelligence.

View arXiv pageView PDFProject pageGitHub19Add to collection

Get this paper in your agent:

hf papers read 2609\.39601

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### GroundingPI/GroundingPI Image-Text-to-Text• 5B• Updated9 days ago • 211 • 9

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.39601 in a dataset README.md to link it from this page.

Spaces citing this paper3

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models

Hugging Face Daily Papers

ACE-Brain-0.6 通过约150万条可验证视觉定位数据微调 Qwen3.5-9B，在七个公开定位基准中的四个上超越 GPT-6 Astra，展示了扩展多样化定位监督数据可显著提升开放多模态模型的空间感知与具身操作能力。

Jag: Direct Box Prediction for Efficient Visual Grounding

Hugging Face Daily Papers

介绍 Jag，一个 0.8B 视觉定位模型，通过单次前向传播直接预测边界框而非自回归生成坐标，在保持与 LocateAnything 相当精度的同时，速度提升 3.62 倍、峰值 GPU 显存降低 73.88%，并开源了模型权重、代码与在线演示。