Xpeng is pushing deeper into physical AI. The company says researchers affiliated with Xpeng, Peking University, The University of Hong Kong and other institutions have released GroundingPI, an open-source visual perception foundation model, with code and weights publicly available.
The roughly 4-billion-parameter model scored an average of 73.68 across 34 perception benchmarks in the research team's evaluation, ahead of OpenAI's GPT-6 Astra at 71.54. CnEVPost notes the score weights 11 capability categories equally, and that the model is trained on perception data without robot action labels.
In simulation, Xpeng says a system using GroundingPI as its visual foundation beat comparison models trained on 75% of demonstration data while using only 50%. CnEVPost reports the research is at the scientific prototype stage and is aimed at the fast «System 1» layer of embodied AI rather than replacing large planning models.