XPeng and partners release GroundingPI, a 4B open perception model they say beats GPT-6 Astra on 34 benchmarks
The researchers report an aggregate 73.68 versus 71.54 for GPT-6 Astra across 34 perception benchmarks, and similar simulated robot-manipulation success with half the demonstrations; results are the team's own and not independently checked.
Why it matters
Robots fail often because they misread what they see, so a small open model focused on perception could matter more to robot makers than another general chatbot. GroundingPI was built by researchers affiliated with XPeng, Peking University and the University of Hong Kong. In simulated manipulation it reached a 28.75% success rate using half the demonstrations, against 27.75% for the strongest rival using 75%. The gains are narrow and self-reported; watch for independent replication and real-robot tests.
Key facts
- About 4 billion parameters, open source
- Aggregate 73.68 vs 71.54 for GPT-6 Astra
- 34 perception benchmarks, 44 comparison models
- 28.75% simulated success with half the demonstrations
- With Peking University and University of Hong Kong researchers