XPENG-backed GroundingPI is a 4B open model that teaches robots where things are

GroundingPI was built by researchers affiliated with XPENG, Peking University, the University of Hong Kong and other institutions[1]. Its focus is a narrow but basic skill for robots and cars: identifying objects and pinning down exactly where they are. Asked for "the handle of the cup on the right", a system has to find the right cup and then the handle's precise position, and GroundingPI answers with spatial coordinates such as points and bounding boxes[1].

It was trained across 11 perception categories, including object detection, dense object localisation, referring-expression grounding, OCR, document layout, user-interface element recognition and robotic spatial localisation, and notably without robot action labels[1]. The idea is to give downstream action-learning systems a stronger visual foundation[1].

Across 34 public perception benchmarks spanning those 11 categories, GroundingPI scored 73.68 on an equally weighted aggregate, the highest of 44 baselines and ahead of GPT-6 Astra's 71.54[1]. It did especially well on scenes with many objects or small targets[1]. XPENG is careful to say this is a perception comparison and not a broader verdict on general intelligence or reasoning[1].

The team also plugged GroundingPI in as the visual backbone for robot and driving models[1]:

• Less training data. In RoboCasa-GR1 simulations, systems using GroundingPI hit higher task success with 50% of the demonstration data than comparison models trained on 75%.
• Better in unfamiliar settings. In RoboTwin 2.0, it outperformed mainstream visual backbones across four out-of-distribution settings.
• Driving. On the nuScenes dataset it reached an average open-loop trajectory prediction error of 0.296 metres.

These results come from simulation and offline evaluation, and XPENG says they don't establish readiness for real-world deployment[1].

Collecting robot demonstrations is slow and expensive, so a vision model that cuts the data needed to learn a task is useful well beyond XPENG. GroundingPI is also a bet that a small specialist perception model, paired with a general model for reasoning and planning, can beat asking one giant model to do everything[1]. XPENG frames it as academic research, not a product announcement for its cars or humanoid robots[1].