XPENG's 4B GroundingPI model beats GPT-6 Astra across 34 grounding benchmarks, boosts robot manipulation
XPENG-AI · hf · 2026-10-13
XPENG AI, with Peking University, HKU, Berkeley, Princeton and others, released GroundingPI, a 4B grounding foundation model aimed at physical intelligence. Key points:
- Method: encodes points/boxes as quantized coordinates in a shared vocabulary; trained with multimodal + spatial pretraining, SFT, and GRPO-based RL, using public datasets plus dedicated data engines.
- Results: new SOTA across 34 grounding benchmarks spanning 11 perceptual capabilities against 44 baselines, averaging 73.68% — above the larger GPT-6 Astra (71.54%).
- Downstream: as a visual backbone for robotic manipulation and autonomous driving, it beats all mainstream backbones in all four RoboTwin 2.0 OOD settings (up to 24.8% relative gain), matches baselines trained on 75% of RoboCasa-GR1 demos while using only 50%, and reaches 0.296 m average open-loop L2 error on nuScenes.
- Takeaway: dense grounding substantially benefits both robotics and driving, and OCR data acts as a catalyst for perceptual learning — dedicated perceptual pretraining is a promising route for physical-intelligence foundation models.
More from Embodied
- Humanoid hand wraps wontons as Nature Will releases Natura-1, a foundation model trained on 1M hours of human data — CyberRobooo · 2026-10-13
- Natura-1 runs multi-minute autonomous dexterous hand workflows with a single policy — stingning · 2026-10-13
- Natura-1 demo: a single-take 30-minute recording of everyday human-robot interaction — stingning · 2026-10-13
- Collov Labs launches NewEyes AI: camera glasses plus open-ear earphones give agents eyes — CodeByPoonam · 2026-10-13
- Surface Ultra unboxed: NVIDIA RTX Spark N1X chip, 20 cores, 64GB unified memory — TinfoilTricorn · 2026-10-13
- Replicating laundry real2sim in MuJoCo and training RL WBC in 3 hours with Opus 5.5 — ZeYanjie · 2026-10-13