Tuesday, 13 October 2026 · 6 min read
GroundingPI improves precise visual grounding for physical intelligence tasks
GroundingPI is a 4B-parameter foundation model that generates quantized points and boxes for precise visual grounding, achieving state-of-the-art results on 34 benchmarks and improving robotic manipulation and autonomous driving performance.
Based on the paper GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives by Qize Yu, Lianrui Fan, Boyu Chen, Jiaqi Liang and 22 others
Challenges in precise visual grounding for physical intelligence
Visual grounding links language to specific objects and locations in images, which is crucial for physical intelligence tasks like robotic manipulation and autonomous driving. Precise grounding must identify the exact target and its location, even in cluttered scenes or for tiny objects, and operate fast enough for closed-loop control.
Current vision-language-action (VLA) and world-action models (WAMs) rely on general-purpose vision-language or video-generation backbones that are not optimized for the precise spatial grounding needed for physical interaction. This leads to perception bottlenecks that limit downstream action learning.
Evaluations show that existing models, including large ones like GPT-6 Astra, struggle in dense scenes and with tiny objects. This motivates developing a dedicated grounding foundation model that can provide precise, broad, and fast perception as a native capability to support physical intelligence.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short