XPENG and Academic Partners Release GroundingPI, a Pretrained Model for Embodied Perception, Outperforming GPT-6 Astra in Overall Perception Benchmarks. $XPEV https://x.com/i/article/2117921113757847552
XPENG and Academic Partners Release GroundingPI, a Pretrained Model for Embodied Perception
XPENG and Academic Partners Release GroundingPI, a Pretrained Model for Embodied Perception, Outperforming GPT-6 Astra in Overall Perception Benchmarks
XPENG, in collaboration with research teams from Peking University and The University of Hong Kong, and other leading academic institutions, has announced GroundingPI, a research breakthrough in foundation models for embodied perception.
GroundingPI introduces a 'perception-oriented' approach. With approximately 4B parameters, the model ranks first in aggregate evaluations across 34 public perception benchmarks, demonstrating greater data efficiency and stronger generalization capabilities in robotic simulation tasks.
Website / Paper / Code / Model
Physical AI and embodied intelligence are emerging as key directions in the development of the AI industry. However, existing general-purpose foundation models are primarily designed for digital applications such as office productivity, conversation, and programming. When directly applied to physical environments such as autonomous driving and robotics, they face inherent gaps in adaptation. The industry is currently exploring two main technical approaches. The first involves fine-tuning and adapting general foundation models for embodied training. The second focuses on developing native embodied models, which are designed for embodied tasks from the outset and trained exclusively on robot-related data. Both approaches are evolving rapidly, with model development cycles typically lasting around six months.
GroundingPI takes a different approach. It proposes that accurate perception of spatial positions and surrounding objects is a fundamental prerequisite for embodied intelligence. Perceptual representations are critical to Physical AI. Without accurately locating an object in an image, even extensive knowledge of actions cannot be effectively translated into real-world execution.
Based on this principle, GroundingPI does not use robot data annotated with joint-level action labels during training. Instead, it is trained entirely on perception-only data without action labels. Its training data covers 11 categories of perceptual capabilities, including general object detection, dense small-object detection, referring expression grounding, OCR, document layout understanding, UI element recognition, and robotic spatial localization. The goal is to first enable the model to understand 'what an object is and where it is located in an image'. For example, when asked to locate 'the handle of the cup on the right', the model must identify the precise locations of both the cup and its handle using points or bounding boxes. Simply answering 'there is a cup in the image' is not sufficient. The objects referred to in language must be accurately grounded in specific regions of the visual scene.
The project name reflects this philosophy: 'Grounding' represents its core perceptual capability of visual localization, while 'PI' stands for Physical Intelligence.
The research team evaluated GroundingPI against OpenAI's frontier model GPT-6 Astra using the same set of 34 perception benchmarks. With equal weighting across the 11 capability categories, GroundingPI achieved an overall score of 73.68, compared with 71.54 for Astra.
Embodied intelligence systems are generally divided into two components: System 1 and System 2. System 1 is the fast system, analogous to the human cerebellum. It is responsible for high-speed, low-latency environmental perception and action execution. System 2 is the slower system, analogous to the human brain, responsible for task understanding and planning—areas where general-purpose foundation models have traditionally excelled. Currently, the industry generally relies on a single general model to serve both systems, distinguishing their functions primarily by adjusting model parameters. XPENG researchers emphasize that model training should be guided by First Principle, and that directly validating the success rate of action execution in the fast system provides the most straightforward approach to evaluating training effectiveness. GroundingPI strengthens perceptual capabilities upstream while retaining the native action learner downstream. Rather than replacing the general-purpose intelligence of System 2, it provides a perception-native foundation for System 1 and offers methodological guidance for pretraining fast-system models.
The research is currently at the research-prototype stage, while its experimental design already points toward a clear path for real-world deployment in embodied intelligence. Current simulation results address two critical challenges in practical deployment:
First, significantly reducing the cost of human demonstrations. Robots typically require hundreds or even thousands of manually collected demonstration samples to learn each new task. Systems built on GroundingPI as their visual foundation achieve higher task success rates with just 50% of the demonstration data required by comparison models using 75% of their data. This improvement comes from GroundingPI directly taking on the task of learning 'where objects are'—a function traditionally learned within Vision-Language-Action (VLA) models. In practical deployment, this means that significantly fewer human demonstrations may be required to achieve the same target success rate.
Second, delivering stronger generalization across different environments. One of the major challenges in real-world robotics is that a robot may fail to perform a task it has already learned when moved to a different workspace or presented with a different object. GroundingPI's perception capabilities enable robots to maintai robust performance despite variations in material properties, object categories, containers, backgrounds, and other environmental factors, minimizing the impact of changing conditions on execution performance. Experimental results demonstrate that in high-density industrial scenarios, such as bulk material sorting and bin inventory counting, systems powered by GroundingPI can accurately classify and pick similar workpieces while quickly adapting to different object shapes and configurations. These results also provide experimental support for the possibility of reusing a single set of model weights across multiple workstations, reducing the need to retrain models for each individual deployment environment.
GroundingPI represents an important building block in XPENG Group's Physical AI ecosystem, introducing a new approach to model pretraining. Its demands for data and computing resources make this technical direction relatively rare within the industry. The project's source code and model weights have been open-sourced. XPENG welcomes researchers and industry partners to explore and validate this approach together.