XPENG open-sources GroundingPI, a 4B perception base model for physical AI

XPENG researchers, working with academic partners including Peking University and the University of Hong Kong, have introduced GroundingPI, an open-source visual perception model positioned as a grounding foundation model for physical AI. The paper is public, and weights have been released on GitHub and Hugging Face.

GroundingPI is a four-billion-parameter model whose core job is to locate and label objects in images, mapping language descriptions onto specific regions of a scene. In the framing used by the researchers, it is meant to be a robot's eyes, supplying the spatial understanding that grasping, navigation and manipulation depend on.

The team evaluated it against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, reporting an average score of 73.68% against 71.54% for the larger GPT-6 Astra. That comparison comes from the researchers' own evaluation setup rather than a third-party rerun, which is worth keeping in mind.

The more meaningful results are downstream. The paper reports gains on robotic manipulation and autonomous driving, including manipulation tasks where better performance was reached with roughly half the training data used by competing systems. That points to reduced data requirements rather than a purely cosmetic benchmark win.

Methodologically, the work uses a staged training recipe: general grounding capability first, then transfer toward physical intelligence scenarios. This “perception foundation plus downstream adaptation” approach speaks directly to the shortage of high-quality labelled data that most robotics teams face.

The boundaries deserve stating. Reporting notes that XPENG said the open-source release is not an announcement that the model is coming to its production vehicles, making GroundingPI a capability release for researchers and developers rather than a product launch.

For a four-billion-parameter perception model, the real signal is not the benchmark lead but that XPENG chose to hand out a robot's eyes as open weights. Carmakers and embodied-AI companies have long treated the perception stack as core intellectual property, and open-sourcing a reusable grounding layer changes how that layer competes.

Two things to watch next: whether the community can reproduce the manipulation gains on real robot platforms, and whether open perception models like this become the default middle layer for carmakers and robotics firms. If they do, differentiation shifts toward data and post-training rather than training a grounding model from scratch.