Researchers affiliated with XPENG, Peking University, The University of Hong Kong, and other academic institutions have introduced GroundingPI, an open-source visual perception foundation model designed for Physical AI applications such as robotics and autonomous driving. The research explores how specialized visual understanding capabilities can help AI systems accurately identify objects and understand spatial positions in physical environments.
Model Specifications and Benchmark Performance
GroundingPI features approximately 4 billion parameters. In evaluations conducted by the research team across 34 public perception benchmarks spanning 11 capability categories, the model achieved an average aggregate score of 73.68. The benchmark evaluation encompassed tasks such as object detection, dense object localization, referring expression grounding, optical character recognition, document layout understanding, user interface element recognition, and robotic spatial localization.
Simulation Results and Embodied Intelligence
The research team evaluated GroundingPI as a visual backbone for downstream robotic manipulation and autonomous driving tasks using simulation environments. In RoboCasa-GR1 simulation experiments, robotic systems utilizing GroundingPI achieved higher task success rates using 50% of the available demonstration data compared to baseline models trained on 75%. Additionally, in RoboTwin 2.0 evaluations across four out-of-distribution settings, the model demonstrated adaptability to changing environments. When integrated for autonomous driving perception using the nuScenes dataset, GroundingPI achieved an average open-loop trajectory prediction error of 0.296 metres.
Open-Source Availability and Scope
The research project is exploratory and based on simulation and offline evaluations, without establishing readiness for commercial deployment or integration into XPENG’s commercial products. To encourage ongoing collaboration, the research team has made the research paper, model weights, and supporting source code publicly accessible to the AI research community.
Practical Takeaway
For European automotive and technology observers, GroundingPI highlights ongoing academic and industry research into foundational visual perception for autonomous driving and robotics, though the open-source release remains strictly an exploratory research tool rather than an immediate commercial vehicle feature.
Source: Original report