Can you guess what this image depicts?These are litter map predictions generated using recursive deep neural networks.

Our colleagues at The Technical University of Cluj-Napoca developed a deep learning model for prediction of litter maps. This model is trained by feeding it with the robot’s trajectory and local litter observations to develop a predicted map. In a future extension of the method, this map will be used by the robot to make decisions on where to go next to discover more litter quickly.

The model is working very well in simple simulated scenarios, so what’s next? Testing it in scenarios closer to real life, until it’s ready to be integrated on the SeaClear2.0 robotic system.

Image Description

Left column: Examples of real maps, where black indicates no litter and white represents a litter object. These images also show the trajectories of the robot on the map.
Middle column: Predicted map in low resolution.
Right column: True map.