Building Row Detection at Werkr Tractors

One of our ongoing projects at Werkr Tractors is developing row detection so our tractors can weed autonomously with high precision. Row following is one of the most critical challenges in automation: even small deviations can damage crops or miss weeds. By

combining smart hardware and machine learning, we’re building a system that can recognize crop rows and guide tractors with centimeter-level accuracy.

Hardware Setup

For data collection, we use a Stereolabs ZED stereo camera, mounted directly on the tractor with a custom 3D-printed bracket. The stereo setup allows us to capture high-resolution images of the crop rows. Alongside the camera, we use an RTK GPS system to provide reliable global positioning for general tractor guidance. However, GPS alone isn’t accurate enough for row-level precision, which is why our machine learning row tracking system steps in to fine-tune the tractor’s path.

Annotation Tool

To train a model that can recognize rows, we first need labeled data. For this, we built an internal annotation tool designed specifically for crop row labeling.

A screenshot of the tool shows a crop row image with six annotated dots, which together form three crop lines. Annotators can load rosbags directly from our cloud (stored in AWS S3), which the tool automatically downloads, extracts, and prepares for annotation.

The tool includes features that make annotation much faster and less error-prone:

• Pause and resume: annotators can stop midway and continue later without losing progress.

• Keyboard shortcuts: quick commands for “previous image,” “undo,” and “next image” make annotation efficient.

• Automatic saving and upload: when an annotation is complete, a CSV file is automatically generated and uploaded back to our cloud.

At the moment, the tool is used internally, but our vision is to scale annotation with:

• Automatic assignment of data to annotators,

• ML-in-the-loop assistance, where the model provides suggestions that humans can quickly refine.

This combination of automation and human oversight will let us scale our dataset efficiently.

Progress on Row Detection

We continue to test our row detection system on a variety of crops (carrots, beets, green onions, lettuce, and more)  to make the model robust across different conditions.

An example GIF of our pipeline shows:

1. Raw images captured in the field,

2. Ground truth labels, where crop rows are annotated as three semantic classes,

3. Predicted segmentations from our ML model, and

4. Fitted lines, where the segmentation is used to estimate the exact position of crop rows.

Although the system is still in its early stages, the qualitative results are promising. With more data, we expect the model to generalize better and handle the natural variability of farms.

Why This Matters

For farmers, accurate row detection means less manual labor, higher weeding precision, and ultimately healthier crops. For us at Werkr Tractors, it’s another step toward delivering practical, field-ready autonomy that reduces costs and increases productivity.

As we continue collecting data and refining our tools, we’re excited to move closer to tractors that can reliably follow rows and weed with minimal human intervention.