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

Key points

• Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.1
• As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving.1
• We systematically analyze GroundingPI's pretraining in scale and data composition.1
• These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.1

Sources (1)

• [1]GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
  Hugging Face Daily Papers · Sep 30, 12:00 AM

  We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary.

  Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

Extractive summary: sentences quoted from the sources.