Robotics and autonomous driving teams keep running into the same wall: vision-language-action (VLA) and world-action models borrow their perception from general-purpose vision-language or video-generation backbones, and those backbones fail in clutter, on tiny objects, and at the speed closed-loop control demands. GroundingPI, released by XPENG AI on September 30, targets exactly that gap: a 4B-parameter visual grounding foundation model, with the paper on arXiv the same day, weights opened October 1, and source code plus deployment and batch-annotation guides following on October 3.
A 4B model topping 34 benchmarks on self-report
The task is deliberately narrow: given an image and an instruction, output points or boxes, with coordinates quantized as integer tokens on a 0-999 grid (the GAM spatial-token protocol). Training proceeds in three stages — multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning — with supervision from public datasets plus dedicated data engines. The official evaluation spans 34 grounding benchmarks, 44 baselines, and 11 perceptual capabilities: GroundingPI averages 73.68%, which the paper self-reports above the larger GPT-6 Astra at 71.54%, claiming a new state of the art. The first author added a positioning note in the Hugging Face paper comments: teams using Grounding DINO can try this as a drop-in upgrade — 4B parameters, with a complete inference stack.
Architecture from off-the-shelf parts, engineering-grade delivery
The three components are all existing open-source pieces: a MoonViT-V2 visual encoder (from Kimi K3), a learnable multimodal projector, and a Qwen3-4B language backbone. The idea is not a new architecture but treating perception as a first-class citizen for dedicated pretraining. Delivery is unusually engineering-grade: vLLM Transformers-backend serving, an OpenAI-compatible API, BF16 precision, a 16,384-token context, and TP=1 on a single card. Officially tested accelerators cover NVIDIA B300, B200, H200, H800, and PPU. Task coverage spans detection, referring boxes, point localization, OCR, document layout, GUI grounding, and visual prompting. The license is Apache 2.0, with Kimi-derived material governed by the Kimi K3 License.
Downstream transfer is the real story
Grounding scores are the appetizer; the main course is downstream transfer. On RoboTwin 2.0, GroundingPI as the visual backbone beats every mainstream backbone evaluated across all four out-of-distribution settings, by up to 24.8% relative to the strongest one. On RoboCasa-GR1, training with 50% of the demonstrations outperforms baselines trained with 75% — a data-efficiency signal worth more than any absolute score. On nuScenes, used as the visual backbone for autonomous driving, it reaches an average open-loop L2 error of 0.296 meters.
A self-graded report card: half cold water, half signal
To be clear: the 73.68%-beats-GPT-6-Astra claim currently comes entirely from the official self-evaluation, cross-checked between two official artifacts (the paper and the GitHub README) but with no independent third-party replication yet — discount accordingly. The direction, however, deserves attention. Using general VLMs as perceptual engines means decent-at-everything, sharp-at-nothing; carving perception out for dedicated pretraining is essentially building a System-1 foundation model for embodied stacks, aligned with the fast/slow dual-system narrative of the world-model camp. One more detail carries more signal than any single score: XPENG has this year shipped TuringViT, the X-AuT in-car speech pruning work, and now GroundingPI — automotive AI teams are shipping foundation-model artifacts, and that trend outlasts any leaderboard. For embodied and autonomous-driving teams, there is now a 4B, vLLM-deployable, permissively licensed perception backbone to try; for general VLM teams, the specialized-perception detour is worth watching before it overtakes.
References: paper at arXiv:2609.39601 (https://arxiv.org/abs/2609.39601); code and weights at https://github.com/groundingpi/GroundingPI