Event

GroundingPI Releases 4B Physical Intelligence Grounding Foundation Model

1 reports · 1 sources

The GroundingPI author team has released GroundingPI, a 4B-parameter grounding foundation model that outputs points and boxes in a shared vocabulary as quantized coordinates, with training combining multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. The author team reports that across 34 grounding benchmarks covering 11 perceptual capabilities and 44 baselines, GroundingPI averages 73.68%, above the larger GPT-6 Astra's 71.54%, establishing a new state of the art. On downstream tasks, the author team reports that GroundingPI surpasses all evaluated mainstream backbones in all four out-of-distribution settings of RoboTwin 2.0, with a maximum improvement of 24.8% over the strongest backbone; on RoboCasa-GR1, GroundingPI trained with 50% of the demonstration data exceeds baselines trained with 75% of the data; and on nuScenes, when used as a vision backbone, it achieves an average open-loop L2 error of 0.296 meters.