GroundingPI 4B Model for Precise Physical Intelligence
September 29, 2026
GroundingPI is a 4B parameter foundation model that generates quantized coordinate points and boxes for precise object localization. It uses multimodal and spatial pretraining combined with GRPO reinforcement learning to outperform 44 baselines across 34 grounding benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can use this for more precise closed-loop control in robotics and vision-language-action tasks.
●
researcherThe use of GRPO for spatial grounding provides a new training paradigm for physical intelligence models.