Detects objects from a natural-language prompt and produces a single pick target suitable for a pick-and-place task. Includes a simple single-image inference script, a batch pipeline, and a pick-task selection pipeline.

• Runs GroundingDINO to detect objects from text prompts.
• Outputs ranked detections and a single recommended pick target.
• Saves annotated images and JSON results for easy demo/review.

• infer.py: single-image inference and visualization.
• batch_infer.py: run inference on a directory of images.
• pick_task_pipeline.py: select a best pick target with JSON output.

1. Install the library:

   pip install groundingdino-py

2. Download weights and config into the project root:

   wget https://huggingface.co/ShilongLiu/GroundingDINO/resolve/main/groundingdino_swint_ogc.pth -O groundingdino_swint_ogc.pth
   wget https://raw.githubusercontent.com/IDEA-Research/GroundingDINO/main/groundingdino/config/GroundingDINO_SwinT_OGC.py -O GroundingDINO_SwinT_OGC.py

3. Run inference:

   python infer.py --image data/image.jpg --prompt "person"

python infer.py --image data/image.jpg --prompt "person"

python batch_infer.py --input data/ --prompt "red mug" --save-annotated --output-dir outputs

python pick_task_pipeline.py --image data/image.jpg --prompt "red mug" --save-annotated

• infer.py saves output.jpg with detections.
• batch_infer.py saves outputs/results.json and optional annotated images.
• pick_task_pipeline.py saves pick_result.json and optional annotated image.

• Place groundingdino_swint_ogc.pth and GroundingDINO_SwinT_OGC.py in the project root.
• For best results, use clear prompts like “red mug on table”.

• Warehouse picking (identify a specific SKU from a shelf)
• Assistive robotics (find objects by voice prompt)
• Manufacturing QA (detect missing or wrong parts)
• Retail analytics (count or locate products on shelves)

• ROS 2 node for real robot integration
• Depth-aware picking using RGB-D
• Multi-object task planning (ordered pick list)
• Web demo UI with live camera input