GroundingPI 4B tops GPT-6 Astra across 34 grounding benchmarks

A 4B-parameter grounding model called GroundingPI averages 73.68% across 34 grounding benchmarks in a new arxiv preprint, edging the larger GPT-6 Astra at 71.54%. The authors describe it as "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 claim that will matter most for robotics teams is downstream. On RoboTwin 2.0, GroundingPI "outperforms every mainstream backbone we evaluate in all four out-of-distribution settings, by up to 24.8% relative to the strongest backbone," the paper reports. On RoboCasa-GR1, a version trained on 50% of the demonstrations beats baselines trained with 75%. Used as a visual backbone on nuScenes, it attains an average open-loop L2 error of 0.296 m.

The abstract names 44 baselines across 11 perceptual capabilities but publishes no per-benchmark numbers, and the 24.8% figure is the top of a relative range in out-of-distribution settings, not an across-the-board average. The authors argue the result "support[s] grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence," and flag "OCR's potential as a catalyst for perceptual learning" in their data analysis.