Xpeng open-sources GroundingPI, edges GPT-6 Astra on perception

Xpeng's new 4-billion-parameter perception model, GroundingPI, scored 73.68 across a 34-benchmark suite, narrowly beating OpenAI's GPT-6 Astra at 71.54 on the same tests, CnEVPost reported. The Chinese automaker built the model with research teams from Peking University and the University of Hong Kong, and has posted code and weights on GitHub.

The name splices "Grounding" (visual localization) and "PI" for Physical Intelligence. In Xpeng's framing, "existing general-purpose large models face an adaptation gap when applied to physical scenarios" like autonomous driving and robotics, so the model trains without robot action labels, leaning on pure perception data for detection, text recognition and spatial localization. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, drawing on public datasets and dedicated data engines.

The downstream claim Xpeng is leading with: a system using GroundingPI as its visual backbone reportedly matches comparison models' success rate on only 50% of the demonstration data, versus 75% for baselines. Material sorting and bin counting are the early use cases, with Xpeng flagging autonomous driving and humanoid robots as the longer target.

The benchmark lead comes from Xpeng's own evaluation on a suite it assembled, with no third-party rerun yet, and the project is still described as a scientific prototype. It lands into a busy stretch of Chinese embodied-AI releases, with our tracker counting 296 China AI stories in the last 90 days alongside Baidu's Apollo World robotaxi trainer landing the same day.