GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
GroundingPI is a 4B model built to make object grounding precise enough for robotics and driving loops.
The paper says it outputs points and boxes as quantized coordinates in a shared vocabulary. It reports state-of-the-art results across 34 grounding benchmarks, averaging 73.68% and beating GPT-6 Astra’s 71.54%. Used as a visual backbone, it improves robotic manipulation and autonomous-driving results, including out-of-distribution RoboTwin 2.0 tests and nuScenes open-loop error. The authors argue that dense grounding and dedicated perceptual pretraining are useful foundations for physical intelligence.
HF Daily Papers' note
The paper says it outputs points and boxes as quantized coordinates in a shared vocabulary. It reports state-of-the-art results across 34 grounding benchmarks, averaging 73.68% and beating GPT-6 Astra’s 71.54%. Used as a visual backbone, it improves robotic manipulation and autonomous-driving results, including out-of-distribution RoboTwin 2.0 tests and nuScenes open-loop error. The authors argue that dense grounding and dedicated perceptual pretraining are useful foundations for physical intelligence.
HF Daily Papers' note
score 6