XPeng open-sources GroundingPI, a 4B grounding model beating larger models across 34 benchmarks
_akhaliq · x · 2026-10-14
XPeng's AI team released GroundingPI on Hugging Face: a 4B grounding foundation model for physical intelligence that uses precise points and boxes as visual primitives. It beats 44 baselines across 34 grounding benchmarks—including larger VLA and world models like Qwen3-VL, PaliGemma-3 and Wan2.2—and leads on autonomous driving (nuScenes) and robotics (RoboCasa/RoboTwin) tasks, with a roadmap of basic perception first, action learning next.
Related event: XPeng Open-Sources GroundingPI, a 4B Grounding Model Beating Larger Rivals(2 posts)→
More from Embodied
- Robotics researcher to students: master coordinate frames, SE(3), and kinematics before deep learning — YuXiang_IRVL · 2026-10-14
- Argon reads the Spot SDK docs by itself and teaches Boston Dynamics' robot dog to trash a can — ki_ki_ki1 · 2026-10-14
- Google's Gemini 4 Argon teaches humanoid robot cornhole zero-shot from a single video, tops ERQA at 87% — bousmalis · 2026-10-14
- Apple HomeView launch video mocked for device's power cord sitting mid-kitchen-island — steipete · 2026-10-14
- Waymo launches robotaxi service to and from Nashville airport (BNA) — reed · 2026-10-14
- Robot learns cornhole by building its own sim, succeeds on real hardware in 6 trials — ki_ki_ki1 · 2026-10-14