XPENG's 4B grounding foundation model GroundingPI hits SOTA across 34 benchmarks
CyberRobooo · x · 2026-11-06
XPENG, with Peking University, HKU, Tsinghua, Princeton and Berkeley, released GroundingPI — a 4B-parameter visual grounding foundation model toward physical intelligence. Built on visual primitives with a staged training recipe, it achieves SOTA across 34 benchmarks and briefly ranked #1 on Hugging Face Daily Papers. It helps humanoid robots understand surroundings, identify objects, interpret pointing gestures and text, with strong ID/OOD transfer and better action-data efficiency in autonomous driving and robotic manipulation evals. Weights are open on Hugging Face and ModelScope, with code, evals, paper and demo released. Seen as the perception base for XPENG's IRON humanoid.
Related event: XPeng Unveils GroundingPI Visual Grounding Model(2 posts)→
More from Embodied
- AI2 Releases ToyBlocks, an Open Simulated Tabletop for Robot Learning Experiments — allen_ai · 2026-11-07
- Somnia Lab is building Model L, an intimacy-focused humanoid robot — 5SecondsNaVzdoX · 2026-11-07
- Scaling H2R workshop at CoRL 2026 unveils speakers, submissions due in 5 days — mangahomanga · 2026-11-07
- Jim Fan to keynote CoRL on Nov 9, teasing new robotics work — DrJimFan · 2026-11-07
- Hermes-Agent hacks Echo Show 5: screen is just a web page you can customize — alexcovo_eth · 2026-11-07
- RoboPapers to host author mixer at CoRL conference — chris_j_paxton · 2026-11-06