GroundingPI Sets New Benchmark for Physical Intelligence with 73.68% Accuracy Across 34 Grounding Tasks

—

Executive Summary & Core Discovery

GroundingPI, a 4B grounding foundation model, sets a new benchmark for physical intelligence with visual primitives. The model generates points and boxes as quantized coordinates in a shared vocabulary, combining multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI achieves an average accuracy of 73.68%, surpassing the larger GPT-6 Astra by 2.14%. As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving, outperforming every mainstream backbone evaluated in all four out-of-distribution settings on RoboTwin 2.0 by up to 24.8% relative to the strongest backbone. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m.

Architectural Background & Theoretical Foundations

GroundingPI is a 4B grounding foundation model designed to address the limitations of current vision-language-action (VLA) and world-action models (WAMs). These models rely on general-purpose vision-language and video-generation backbones, which struggle with precise grounding in cluttered environments and for tiny objects. GroundingPI generates points and boxes as quantized coordinates in a shared vocabulary, ensuring fast and accurate grounding for closed-loop control. The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

Benchmark Architecture & Evaluation Methodology

GroundingPI was evaluated against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities. The evaluation methodology includes systematic analysis of the model’s pretraining in scale and data composition. Downstream autonomous driving and robotic manipulation improve as the pretraining is scaled. The data recipe across these 11 perceptual capabilities shows dense grounding’s substantial benefits for both, and OCR’s potential as a catalyst for perceptual learning.

Empirical Findings & Granular Performance Analysis

GroundingPI achieves an average accuracy of 73.68% across 34 benchmarks, outperforming GPT-6 Astra by 2.14%. As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. On RoboTwin 2.0, it outperforms every mainstream backbone evaluated in all four out-of-distribution settings by up to 24.8% relative to the strongest backbone. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m.

In-Depth Analysis of Core Limitations & Vulnerabilities

GroundingPI addresses several core limitations identified in current vision-language-action (VLA) and world-action models (WAMs). The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. GroundingPI’s success demonstrates the critical role of dedicated perceptual pretraining in foundation models for physical intelligence.

1. Precise Grounding in Cluttered Environments

GroundingPI generates points and boxes as quantized coordinates in a shared vocabulary, ensuring fast and accurate grounding for closed-loop control. The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

2. Performance in Robotic Manipulation Tasks

GroundingPI improves performance on robotic manipulation tasks by up to 24.8% relative to the strongest backbone. The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

3. Performance in Autonomous Driving Scenarios

GroundingPI attains an average open-loop L2 error of 0.296 m, indicating better performance in autonomous driving tasks. The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

4. Dedicated Perceptual Pretraining

GroundingPI’s success demonstrates the critical role of dedicated perceptual pretraining in foundation models for physical intelligence. The model’s architecture combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines.

Enterprise Production & Red-Teaming Blueprint

GroundingPI’s success in setting a new benchmark for physical intelligence with visual primitives highlights the importance of dedicated perceptual pretraining in foundation models. As an automated evaluation layer, JudgeMyAI can help enterprises assess and improve their grounding foundation models by providing comprehensive benchmarking and performance analysis. By leveraging GroundingPI’s architecture and evaluation methodology, enterprises can enhance their visual backbones for robotic manipulation and autonomous driving, ensuring precise grounding in cluttered environments and for tiny objects.

Technical FAQ (Granular & Non-Repetitive)