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

arXiv · Qize Yu, Lianrui Fan, Boyu Chen, Jiaqi Liang, Xini Ding, Yue Chen, Zetian Song, Yuran Wang, Yi Zou, Kaixuan Wang, Tianxing Chen, Wenxuan Song, Bohan Zhou, Mingleyang Li, Siqiao Huang, Yuqi Ye, Caigao Jiang, Wei Wei, Ruihai Wu, Hang Zhang, Yixiao Ge, Shuchang Zhou, Shilong Liu, Xianming Liu, Ping Luo, Shiyu Huang · 2026-09-30

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

GroundingPI evaluates a 4B visual-grounding backbone on grounding, manipulation and open-loop driving benchmarks.

What it does

GroundingPI emits quantized points and boxes in a shared vocabulary. Multimodal/spatial pretraining, supervised fine-tuning and GRPO use public data and dedicated data engines. Across 34 grounding benchmarks and 44 baselines it averages 73.68%, versus the larger GPT-6 Astra at 71.54%.

Why it matters

As a visual backbone, it improves the tested downstream systems: up to 24.8% relative gain over the strongest evaluated backbone on RoboTwin 2.0, and a RoboCasa-GR1 comparison using 50% versus 75% of demonstrations. The nuScenes result is 0.296 m average open-loop L2 error.

How it applies

Compare the grounding backbone on target scenes, including small objects and clutter. For a robot or driving system, add closed-loop control, target-hardware latency and safety evaluation; quantized coordinates and open-loop error do not certify those behaviors.

The short version

Separate better perception from proof of safe control.

My take

The useful adoption question is whether perception errors fall on the target scenes while the full controller remains stable and timely. A simulation score is a starting point for that evaluation.

How it connects

• Perceptual backbone versus control policy
• Relative gains and demonstration budgets
• Open-loop versus closed-loop evaluation

Bottom line

Test target-scene grounding first, then validate the complete physical loop.

Takeaways

• The 73.68% aggregate covers 34 grounding benchmarks.
• The RoboTwin improvement is relative to the strongest tested backbone.
• RoboCasa compares different demonstration fractions under its experimental setup.
• nuScenes measures open-loop error rather than deployed driving safety.
• Exact coordinate representation does not imply error-free physical localization.

Topics: visual grounding, physical intelligence, robotics, autonomous driving, foundation models