Found first: a primary source the press has not covered yet.
A 4B-parameter grounding model from Peking University, HKU, UC Berkeley, Princeton, NUS, and Tsinghua scores 73.68% average across 34 grounding benchmarks, outpacing GPT-6 Astra's 71.54%. The GroundingPI paper, submitted to arXiv on September 30, 2026, also reports up to 24.8% relative improvement over the strongest robotics baseline on out-of-distribution tasks in RoboTwin 2.0.
What the source says
GroundingPI, led by Qize Yu and 25 co-authors across six institutions, generates points and bounding boxes as quantized coordinates unified in a single shared vocabulary. It was trained through multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO on public datasets augmented by dedicated data engines. Evaluation covered 44 baselines across 34 benchmarks spanning 11 perceptual capability categories, with GroundingPI averaging 73.68% against GPT-6 Astra's 71.54%. On RoboTwin 2.0, the model improved over the strongest evaluated backbone by up to 24.8% across four out-of-distribution settings. On RoboCasa-GR1, it exceeded baselines trained on 75% of demonstrations while using only 50%, and it was also evaluated on autonomous driving via nuScenes, reaching 0.296 m average open-loop L2 error. Model weights, code, and a project page are publicly available.
Why it matters
Grounding, localizing objects and regions from language, is a recurring bottleneck in vision-language-action pipelines for robotics. This result puts a 4B open model ahead of a frontier closed system on a 34-benchmark standardized suite, giving robotics teams an auditable, deployable alternative to API-dependent grounding. The RoboCasa-GR1 data-efficiency finding matters separately: teams with constrained demonstration budgets may get more from GroundingPI than from models trained on their full datasets. The nuScenes result suggests the grounding representations transfer beyond manipulation to autonomous driving, though those numbers stand on their own and the paper does not claim equivalence across task types.