XPeng with Peking and Hong Kong universities releases GroundingPI on November 4, topping GPT-6 Astra on perception benchmarks
GroundingPI uses roughly 4 billion parameters with a perception-first approach, ranking first across 34 public perception benchmarks with 73.68 versus GPT-6 Astra at 71.54.
On November 4 XPeng Group together with research teams from Peking University, the University of Hong Kong and others published research results on the embodied perception foundation model GroundingPI. The company said the work proposes a perception-first technical route, ranking first in a combined evaluation across 34 public perception benchmarks at roughly 4 billion parameter scale, and showing higher data processing efficiency and stronger scene generalisation in robot simulation tasks.
The team compared GroundingPI with OpenAI frontier model GPT-6 Astra on the same set of 34 perception benchmarks, with equal weighting across 11 capability categories: GroundingPI scored 73.68 against 71.54 for Astra. GroundingPI argues that the primary precondition for embodied intelligence is accurately perceiving the spatial position of the robot itself and surrounding objects, and that perception representations are essential to physical AI, since a model that cannot localise targets in the frame cannot make action knowledge effective no matter how much it has learned.
On that basis GroundingPI uses no joint action labels in training, relying entirely on pure perception data without action labels, covering 11 perception capabilities including general detection, dense small objects, referring expression localisation, text recognition, document layout, interface elements and robot spatial localisation. The goal is to let the model first understand what an object is and where it sits in the frame. This route choice reflects one branch of the technical divergence in embodied intelligence, from action-data-driven toward perception-representation-first.