MENU

GroundingPI: XPENG’s 4B-parameter model for embodied perception

Gasgoo China
Overview
XPENG, Peking University, and the University of Hong Kong have released GroundingPI, a 4-billion-parameter foundation model for embodied perception. The model outperformed GPT-6 Astra in a 34-benchmark evaluation, scoring 73.68 points compared to 71.54. This development reduces manual teaching costs by achieving higher success rates with significantly less demonstration data.
In Depth

Key Findings

GroundingPI has demonstrated superior performance in embodied perception benchmarks, scoring 73.68 points across 34 public evaluations. This result surpasses the 71.54 points achieved by OpenAI’s GPT-6 Astra. In simulations, systems utilizing GroundingPI achieved a higher success rate using only 50% of demonstration data, whereas comparison models required 75% of the data to reach their results.

Technical Details

The model consists of roughly 4 billion parameters and employs a “perception-first” architecture. Unlike traditional models, GroundingPI avoids robot data labeled with joint actions during training. Instead, it utilizes pure perception data across 11 specific capabilities, including:

  • General detection
  • Dense small-target recognition
  • Referential localization

This approach enhances visual grounding, allowing the model to identify both the identity and the precise spatial location of objects within a frame. Architecturally, GroundingPI is designed as a perception-native base for the “fast system” (System 1 of embodiment), rather than serving as a replacement for the general capabilities of the “slow system” (System 2).

Background and Context

The development addresses the adaptation gap encountered when general large models, which are primarily designed for digital scenarios, are applied to physical environments. By focusing on embodied intelligence and physical AI, the research team aimed to bridge the divide between digital processing and physical world perception.

Significance and Outlook

The reduction in required demonstration data suggests a significant decrease in manual teaching costs and improved generalization across different backgrounds and materials. While the research currently remains in the prototype stage, the team has open-sourced the model weights and relevant code to facilitate further development in embodied perception.

Source: https://autonews.gasgoo.com/articles/icv/xpeng-and-universities-unveil-groundingpi-foundation-model-for-embodied-perception-2118547420962770945

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

  • Your email and selected fields are used only to deliver the newsletter.
  • We never share your information with third parties.
  • You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Takes about a minute · Unsubscribe anytime

Published by Troy-Technical, an independent site run by one engineer with a career in materials development.
About the author / Contact info@troy-technical.jp
Let's share this post !

Author of this article

TOC