GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
GroundingPI, a 4B grounding model, sets a new state of the art across 34 benchmarks and improves robotics and driving.
GroundingPI is a 4B grounding foundation model that predicts points and boxes as quantized coordinates for physical-intelligence tasks. Training mixes multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. Across 34 grounding benchmarks and 44 baselines it averages 73.68%, above GPT-6 Astra at 71.54%. As a visual backbone it leads tested models on RoboTwin 2.0 in all four out-of-distribution settings by up to 24.8% relative, beats higher-data baselines on RoboCasa-GR1 with 50% of demonstrations, and reaches 0.296 m average open-loop L2 on nuScenes.
- 4B model emits points and boxes as quantized shared-vocabulary coordinates.
- Averages 73.68% on 34 benchmarks, above GPT-6 Astra at 71.54%.
- On RoboTwin 2.0 it leads every tested backbone, by up to 24.8% relative.
- nuScenes open-loop L2 error is 0.296 m when used as the visual backbone.
- Scaling pretraining and dense grounding both improve driving and manipulation.
- AI models
- GroundingPIGPT-6 Astra
From the original article
1 min read at huggingface.co
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.
By Qize Yu, Lianrui Fan, Boyu Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.