We have been using Machine Learning, often Deep Learning, extensively as part of SLAM approaches, particularly in mapping. This includes semantic segmentation and instance segmentation, as well as prediction of geometry, e.g. image-based depth prediction and depth completion. Furthermore, understanding uncertainty of predicted quantities allows for robust interfaces with other components running on-board a mobile robot.
As an example of most recent works, we are performing depth completion with a Convolutional Neural Network that also learns to predict depth uncertainty, which is then readily used in probabilistic occupancy mapping. Such systems may allow a robot to safely navigate into space where no depth camera measurements are available (e.g. because the structure is too far away), while maintaining the ability to accurately map the area where actual measurements are available.
Current collaborators: