Xpeng opens GroundingPI, tops GPT-6 Astra at 73.68

Xpeng's framing is that general-purpose large models face an adaptation gap when applied to physical scenarios such as autonomous driving and robotics. GroundingPI is its attempt to close that gap by training without robot action labels, leaning on pure perception data for detection, text recognition and spatial localization. The training recipe combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, drawing on public datasets and dedicated data engines.

The name splices "Grounding" (visual localization) and "PI" for Physical Intelligence, which signals where Xpeng wants this to sit: as a visual backbone rather than a full driving or robot policy. Early use cases named are material sorting and bin counting, with autonomous driving and humanoid robots as the longer target.

The result lands in a busy stretch of Chinese embodied-AI releases. Xpeng's own tracker counts 296 China AI stories in the last 90 days, and Baidu's Apollo World robotaxi trainer landed the same day. Against that backdrop, the open-sourcing of code and weights is likely to be read as a bid for outside testing, though the benchmark lead itself remains self-reported for now.