XPeng open-sourced GroundingPI , a robot perception model that beats GPT-6 Astra

Open source visual perception model

XPeng just open-sourced a 4-billion-parameter visual perception model that outperforms GPT-6 Astra on grounding benchmarks, and lets robots learn physical tasks using half the training data of competing systems.

GroundingPI, developed by XPeng researchers in collaboration with Peking University and the University of Hong Kong, was posted to arXiv on September 30 and hit number 1 on Hugging Face Daily Papers on October 14. 

The code is on GitHub under CC BY 4.0, the weights are on Hugging Face .. ergo ... it's genuinely open.

A 4B model beating a larger frontier system on specialized benchmarks is impressive, as is the data efficiency finding.

What GroundingPI actually does

GroundingPI is not a full vision-language-action model. 

It's a perception backbone, specifically a grounding model. Its job is to link language descriptions to precise spatial locations in images, outputting coordinates rather than semantic labels. When a robot needs to know exactly where "the handle of the blue mug" is in a camera frame, that's a grounding problem. 

General-purpose vision-language models handle this poorly. GroundingPI was built to handle it well.

The architecture combines a MoonViT-V2 visual encoder (from Kimi K3) with a Qwen3-4B language backbone. Coordinates are represented as quantized integers on a 0–999 grid, encoded as special tokens in a shared vocabulary alongside text. That unified representation lets one model handle referring expression grounding, object detection, OCR, document layout, GUI identification, and visual prompting through a single interface.

Training happens in 3 stages: 

1. base multimodal pretraining
2. supervised fine-tuning on coordinate alignment
3. reinforcement post-training using Group Relative Policy Optimization
   

Critically, the model is trained entirely on perception data. There's no labeled robot action data and no joint trajectories, making it easy to adopt.

The numbers

Across 34 benchmarks spanning 11 perceptual capability categories, GroundingPI scores 73.68% average, against 71.54% for GPT-6 Astra, which is a much larger frontier model. 

GroundingPI leads in 8 of the 11 categories, but specialized models outperform it in on-screen GUI detection, crowded scene analysis, and robot-specific tasks.

On robotics benchmarks: 

1. On RoboTwin 2.0, it outperforms every mainstream backbone tested in all 4 out-of-distribution settings, by up to 24.8% over the strongest competitor. Out-of-distribution performance is what matters for deployment, not performance in controlled training conditions. 
2. On RoboCasa-GR1, robots using GroundingPI as their backbone reached equivalent task success with 50% of the demonstration data that competing backbones required at 75%.
   

One caveat: the 34-benchmark evaluation suite was designed by XPeng's team. The ~2-point margin over GPT-6 Astra hasn't yet been independently verified by third-party labs. 

That's standard for fresh arXiv releases, but it's worth remembering as we wait for independent replications.

The strategic play

GroundingPI plugs into the front end of a VLA pipeline. 

Models like OpenVLA and Physical Intelligence's pi0 have their own vision encoders. XPeng's claim, backed by the RoboTwin 2.0 results, is that swapping in GroundingPI as the vision backbone improves performance on novel environments while reducing the amount of demonstration data needed.

Inside XPeng's own stack, GroundingPI serves as the fast system in a fast/slow architecture: real-time spatial perception feeding an action execution layer. The slow system handles high-level task planning. IRON, XPeng's humanoid robot that entered production in September, runs this stack on-device using 3 Turing chips delivering 2,250 TOPS. (GroundingPI's training on pure perception data makes it cleanly separable.) It serves both IRON's manipulation tasks and XPeng's autonomous driving programs from a single model.

Open-sourcing the perception layer while keeping proprietary advantages in action execution and hardware smart. If GroundingPI becomes a standard perception backbone across the VLA ecosystem, XPeng shapes the direction of the field. The IRON hardware and XPeng's action model, however, remain proprietary. The perception frontend then becomes a community asset that also happens to validate XPeng's architectural choices.

Sources

• arXiv — GroundingPI paper 2609
• GitHub — GroundingPI code repo
• GroundingPI — project page
• CnEVPost — XPeng GroundingPI release
• Electrek — XPeng Robotics funding