Robotics field tooling + sensing

ROS2 field sensing and dataset quality for agricultural robotics.

This work connected thermal sensing, GPS metadata, and computer-vision data workflows in a field setting where clean assumptions do not last long.

Back to work

Collection

ROS2 thermal imaging

Thermal imagery published through ROS2 with GPS metadata folded into field collection.

Curation

Representative subsets

Embedding and clustering workflows used to reduce redundant annotation batches.

Context

PerPlant is an agtech startup using computer vision and AI to support precision agriculture. My role sat close to the robotics software layer: getting sensing data collected reliably, preserving useful metadata, and making downstream annotation work less redundant.

Stack / themes

ROS2C++PythonThermal imagingGPS metadataHDBSCAN

Contributions

Integrated a thermal imaging workflow using ROS2 publishers/subscribers in C++ and Python.

Enabled GPS-geotagged thermal image collection for field datasets.

Built a dataset sampling workflow using ROI filtering, detector embeddings, UMAP, HDBSCAN, and grouped splits.

Reduced repeated annotation effort by prioritizing representative and informative image batches.