# XPENG and Partners Release GroundingPI Open-Source Visual Perception Model New 4B-parameter model enhances spatial awareness for robotics and autonomous driving systems. - **Source:** [CleanTechnica](https://cleantechnica.com) - **Published:** 2026-11-06 - **Topic:** [technology](https://brifing.si/c/technology) - **Original article:** - **On Brifing:** https://brifing.si/s/c659493c-f249-49bd-9eb6-47a29773d2db ## New Open-Source Perception Model Researchers from XPENG, Peking University, and The University of Hong Kong introduced GroundingPI, a 4-billion parameter visual perception model. Designed for Physical AI, the model focuses on identifying objects and spatial relationships. The team released the research paper, model weights, and source code publicly to foster collaboration within the global AI community. ## Outperforming General-Purpose Baselines GroundingPI achieved an aggregate score of 73.68 across 34 perception benchmarks, surpassing GPT-6 Astra’s score of 71.54. The evaluation covered 11 capability categories, including object detection and robotic spatial localization. These results highlight the efficiency of specialized perception-focused models compared to larger, general-purpose foundation models in handling precise visual grounding tasks. ## Focus on Physical Intelligence Physical AI requires more than language reasoning; it demands precise visual understanding of the surrounding environment. GroundingPI represents information using spatial coordinates, such as points and bounding boxes, to enable accurate object interaction. This approach allows systems to locate specific targets, like a cup handle, rather than just identifying the object itself. ## Efficiency in Robotic Manipulation In RoboCasa-GR1 simulation experiments, robotic systems using GroundingPI achieved higher success rates with only 50% of the demonstration data compared to models using 75%. This suggests that superior foundational perception can significantly reduce the volume of action-training data required. Such efficiency gains are critical for scaling complex robotic manipulation tasks in real-world environments. ## Adaptability to New Environments Generalization remains a major hurdle in robotics, but GroundingPI showed improved performance across four out-of-distribution settings in RoboTwin 2.0 simulations. By adapting better to unfamiliar backgrounds and environmental conditions, the model demonstrates potential for more robust physical interaction. This adaptability is essential for deploying autonomous systems in diverse, unpredictable real-world settings. ## Autonomous Driving Potential The research team evaluated GroundingPI’s utility for autonomous driving using the nuScenes dataset. When integrated as a visual backbone, the model achieved an average open-loop trajectory prediction error of 0.296 meters. While these results are promising, they are based on simulation and do not confirm readiness for immediate commercial deployment in vehicles. ## Research-Oriented Development Path GroundingPI is an exploratory academic project supported by XPENG to advance Physical AI architectures. The findings emphasize that specialized perception models can effectively complement higher-level reasoning systems. The project does not represent an immediate integration into XPENG’s commercial products, but rather serves as a foundational contribution to the broader robotics research ecosystem. ## Future Implications for Robotics By prioritizing visual grounding, this research provides a new framework for developing adaptable embodied intelligence. Future applications could span humanoid robots, autonomous vehicles, and smart living technologies. As developers continue to iterate on these open-source tools, the industry may see faster progress in creating machines that interact safely and accurately with humans. ## Sources covering this story - https://cleantechnica.com/2026/11/05/xpeng-researchers-and-academic-partners-introduce-groundingpi-an-open-source-visual-perception-model-for-physical-ai/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._