XPeng, Peking University, and HKU Release GroundingPI Pre-trained Model with Top Performance on 34 Perception Benchmarks

IT Home November 4 news, XPeng Group, in collaboration with research teams from universities such as Peking University and the University of Hong Kong, jointly released the research results of the embodied perception base model GroundingPI.

The official statement indicated that the study proposes a “perception-first” technical route, ranking first in a comprehensive evaluation of 34 public perception benchmarks with a scale of approximately 4 billion parameters, and demonstrating higher data processing efficiency and stronger scene generalization capabilities in robot simulation tasks.

The research team compared GroundingPI with OpenAI's cutting-edge model GPT-6 Astra on the same set of 34 perception benchmarks; based on an equal-weighted average of the 11 categories of capabilities mentioned above, GroundingPI scored 73.68, and Astra scored 71.54.

According to the description, GroundingPI proposes that the primary prerequisite for embodied intelligence is the accurate perception of the spatial positions of oneself and surrounding objects; perception representation is crucial for physical AI. If a model cannot accurately locate the position of a target in a scene, no matter how much action knowledge it learns, it cannot be effectively implemented.

Based on this judgment, GroundingPI does not use robot data with joint action labels during the training phase, but instead uses all pure perception data without action labels, covering 11 types of perception capabilities such as general detection, dense small targets, referring expression localization, text recognition, document layout, interface elements, and robot spatial localization, with the goal of letting the model first understand “what the object is and at what position in the screen.”

For example, if we ask the model to find the “handle of that cup on the right,” it needs to mark the specific location of the cup and the handle using points or bounding boxes. Answering “there is a cup in the picture” is not enough; the target in the language must correspond to a specific area in the image.

The project name also originates from this; Grounding corresponds to the core perception capability of visual localization, and PI is the abbreviation for Physical Intelligence.

The official source stated that, This research is currently at the research prototype stage, but its experimental design itself has already indicated the direction of implementation in the embodied field, and the current simulation results have already directly addressed two core issues in quantitative deployment:

• First, significantly reduce teaching costs.For every new task a robot completes, hundreds or thousands of demonstration data points usually need to be manually collected. The system using GroundingPI as the visual base exceeds the success rate of comparison models using 75% of the data with only 50% of the demonstration data. Its mechanism lies in the fact that it directly takes on the task of learning “where the object is” in traditional VLA models. This means that in actual deployment, the amount of manual demonstration required under the same success rate target can be significantly reduced.

• Second, Bring stronger scene generalization.A current pain point in robot deployment is that after switching to another space / changing to another object, the actions that were previously being performed will no longer be performed. The perception capabilities possessed by GroundingPI allow robots to maintain extremely strong stability when dealing with different on-site variables such as material changes, object category changes, container changes, and background changes, and their execution performance is not affected by environmental changes. In high-density scenarios such as loose material sorting and bin counting in simulated environments, GroundingPI's system can classify and grasp similar workpieces and quickly correspond to item shapes. This also provides experimental support for “reusing one set of weights across multiple workstations.”

IT Home attached the relevant links:

• Website: https://groundingpi.github.io/

• Papers: https://arxiv.org/abs/2609.39601

• Code: https://github.com/groundingpi/GroundingPI

• Model: https://huggingface.co/GroundingPI/GroundingPI

Advertising disclosure: External links in this article, including but not limited to hyperlinks, QR (2D) codes and passphrases, provide more information and save readers time in selecting information. Results are for reference only. All IT Home articles containing external links include this disclosure.