Hugging Face has published a research paper introducing GroundingPI, described as a grounding foundation model aimed at physical intelligence. The work argues that precise grounding — specifying which object is the target and where it is, even in clutter and for tiny objects, and fast enough for closed-loop control — is a capability that general-purpose vision-language and video-generation backbones still fail to deliver for vision-language-action (VLA) and world-action models (WAMs).
GroundingPI is a 4B-parameter model that generates points and boxes as quantized coordinates within a shared vocabulary. Its training pipeline combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, drawing supervision from public datasets and dedicated data engines.
The authors report that against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI sets a new state of the art with an average of 73.68%, above the larger GPT-6 Astra at 71.54%. As a downstream visual backbone, it is said to improve robotic manipulation and autonomous driving: on RoboTwin 2.0 it outperforms every mainstream backbone evaluated in all four out-of-distribution settings, by up to 24.8% relative to the strongest backbone; on RoboCasa-GR1 it trained with 50% of demonstrations and still outperformed baselines trained with 75%; and on nuScenes it reached an average open-loop L2 error of 0.296 m.
The paper also reports a systematic analysis of pretraining scale and data composition, finding that downstream autonomous driving and robotic manipulation improve as pretraining is scaled. Analyzing the data recipe across the 11 perceptual capabilities indicated substantial benefits from dense grounding for both tasks, and pointed to OCR as a potential catalyst for perceptual learning.
Code and model weights are described as open source, with a complete inference stack for deployment and additional material at groundingpi.github.io. The authors frame the results as support for grounding as a perceptual foundation and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.