GroundingPI cuts robot vision latency to 5 ms on a single A100

Researchers unveiled GroundingPI, a 4B-parameter model that reframes object grounding as visual-primitive token generation, returning box coordinates in 5 ms on a single A100. It reaches 71.4% mAP-box and 94% pick-place success on a Franka arm at 30 Hz.

• 5 ms inference on A100 versus 45 ms for Meta-Llama-VLA-7B
• 71.4% mAP-box and 78.9% mAP-point on COCO-Grounding
• 94% pick-place success on Franka at 30 Hz, 0.42 J per inference
• Weights, API, ROS bridge and ONNX export released on day one