XPENG Researchers Say 4B GroundingPI Beats GPT-6 Astra on 34 Perception Tests

Researchers affiliated with XPENG, Peking University and the University of Hong Kong release GroundingPI, an open-source 4B-parameter visual perception model. The team reports it scored 73.68 across 34 perception benchmarks, edging GPT-6 Astra at 71.54, and says simulation tests cut robot demonstration data needs. Weights and code are public.

• The team evaluates GroundingPI against 44 baseline models across 34 benchmarks spanning 11 perception capability categories.
• In RoboCasa-GR1 simulations, the researchers report GroundingPI reaches higher success rates with 50% of demonstration data than comparison models using 75%.
• The team reports a 0.296 metre average open-loop trajectory prediction error on the nuScenes autonomous driving dataset.