GroundingPI Debuts: 4B Grounding Foundation Model for Physical Intelligence Outperforms GPT-6 Astra Across 34 Benchmarks

Background and the Problem

In embodied robotics and physical intelligence, precise visual grounding is the non-negotiable prerequisite for closed-loop manipulation. It dictates which object is the designated target and where its geometric manipulation boundaries lie in cluttered, real-world scenes. However, modern vision-language-action (VLA) and world-action models (WAM) universally borrow general-purpose vision-language or video generation backbones. These general models were never optimized for dense spatial precision or tiny parts, resulting in frequent spatial coordinate hallucination and grasp failures under tight closed-loop control frequencies.

Architecture and How It Works

To establish a dedicated foundation for physical perception, researchers unveiled GroundingPI, a specialized 4B grounding foundation model engineered for physical intelligence:

1. Quantized Coordinates in Unified Vocabulary: Abandons traditional separate detector bounding-box heads, representing 2D/3D visual points and bounding boxes directly as quantized coordinate tokens within a unified autoregressive vocabulary;
2. Three-Stage Physical Pre-training Pipeline: Integrates multimodal spatial pretraining, supervised fine-tuning, and GRPO-driven reinforcement learning, supported by specialized spatial data engines;
3. OCR as a Perceptual Catalyst: Mechanistic ablation across 11 perceptual domains revealed that dense text OCR grounding acts as a powerful catalyst for general spatial and geometric reasoning, dramatically elevating tiny-object localization.

Benchmarks and Measured Results

Benchmarked across 34 spatial grounding benchmarks and physical robotics suites, GroundingPI delivers staggering empirical milestones:

• Surpassing Commercial Frontier Models: Outperforming 44 competitive baselines across 11 perceptual capabilities, GroundingPI establishes a new state of the art with an average score of 73.68%, outperforming the larger proprietary GPT-6 Astra (71.54%);
• RoboTwin 2.0 OOD Superiority: When integrated as the visual backbone for robotic manipulation, it beats all mainstream backbones across all four out-of-distribution environments by up to 24.8% relative margin;
• RoboCasa-GR1 Demonstration Efficiency: On household robotic benchmarks, GroundingPI trained with only 50% demonstration data surpassed competitor baselines trained on 75% demonstrations;
• Autonomous Driving nuScenes: Delivered an average open-loop trajectory prediction L2 error of only 0.296 meters, satisfying stringent latency and precision bounds for real-world robotics.

Getting Started for Developers

GroundingPI is open-sourced on GitHub at https://github.com/groundingpi/GroundingPI:

• Quickstart Installation: Clone the repository via git clone https://github.com/groundingpi/GroundingPI to access quantized coordinate decoders and pretrained 4B weights;
• Robotics Pipeline Drop-in Replacement: Developers of robotic arms and autonomous mobile platforms can replace standard CLIP/SigLIP encoders with GroundingPI to immediately resolve spatial blindness in messy workspaces;
• Real-Time Integration: The quantized token architecture enables lightning-fast GPU decoding compatible with ROS2 and NVIDIA Isaac Sim workflows.