College of Science & Engineering
Twin Cities
These researchers are using deep learning-based approaches to solve many relevant computer vision problems that come up in robotics planning tasks. The first such problem is concerned with designing autonomous weed mowing robots. In particular, the group studies the problem of how to accurately detect the weeds on cow pastures and to learn stable navigation strategies for autonomous weed mowing. They also work on some fundamental problems related to robotic planning, such as object detection/segmentation, camera pose estimation, and 3D reconstruction. Training any modern network architecture in a feasible amount of time has become costly even for modern machines with high-end graphics cards. The researchers mainly use the GPU nodes on the Agate cluster, with their large GPU memory footprints, to train the models.
Research by this group was featured on the MSI website in June 2021: Improving Robot Learning Through Visual Forecasting.