New image dataset aims to improve AI-based navigation of PV cleaning robots

Researchers from several universities, including the University of Technology Malaysia and Canada's University of Prince Edward Island, have built a large set of images meant to help robots that clean solar panels "see" better. The goal is to train these cleaning robots to navigate across a panel's surface even when it's dirty, by recognizing the grid lines and borders built into the panel structure rather than relying on a clean surface.

The dataset includes 6,200 images: some clean, some computer-generated with varying levels of grime like bird droppings, water stains, and dust, and some real photos of panels dirtied with sand, pigment, and wastewater. By keeping the panel's underlying line pattern the same across all these versions, the researchers can test whether a robot's navigation software still finds its way when the surface is covered in different kinds and amounts of mess.

This is early-stage research, not a product. The team says the next step is to use this image data to build lightweight AI systems that can judge a cleaning robot's position and direction on a panel, then test those systems on actual robots in real conditions. For a homeowner, this doesn't mean a new cleaning service or device is available now. It's a building block aimed at making automated panel-cleaning robots more reliable in the future, which could eventually make it easier to keep panels free of grime that cuts into their power output.