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

What happened: Researchers introduced GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary, trained with multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO. Key numbers: 4B parameters. Why it matters: The authors report state of the art against 44 baselines and say GroundingPI improves robotic manipulation and autonomous driving as a visual backbone, supporting dedicated perceptual pretraining as a direction for physical intelligence.

Researchers introduced GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary, trained with multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO.

The authors report state of the art against 44 baselines and say GroundingPI improves robotic manipulation and autonomous driving as a visual backbone, supporting dedicated perceptual pretraining as a direction for physical intelligence.