AI Applications on Planetary Rovers
Abstract
The rise in the number of robotic missions to space is paving the way for the use of artificial intelligence and machine learning in the autonomy and augmentation of rover operations. For one, more rovers mean more images, and more images mean more data bandwidth required for downlinking as well as more mental bandwidth for analyzing the images. On the other hand, light-weight, low-powered microrover platforms are being developed to accommodate the drive for planetary exploration. As a result of the mass and power constraints, these microrover platforms will not carry typical navigational instruments like a stereocamera or a laser rangerfinder, relying instead on a single, monocular camera.
The first project in this thesis explores the realm of novelty detection where the goal is to find `new' and `interesting' features such that instead of sending a whole set of images, the algorithm could simply flag any image that contains novel features to prioritize its downlink. This form of data triage allows the science team to redirect its attention to objects that could be of high science value. For this project, a combination of a Convolutional Neural Network (CNN) with a K-means algorithm as a tool for novelty detection is introduced. By leveraging the powerful feature extraction capabilities of a CNN, typical images could be tightly clustered into the number of expected entities within the rover's environment. The distance between the extracted feature vector and the closest cluster centroid is then defined to be its novelty score. As such, a novel image will have a significantly higher distance to the cluster centroids compared to the typical images. This algorithm was trained on images obtained from the Canadian Space Agency's Analogue Terrain Facility and was shown to be effective in capturing the majority of the novel images within the dataset.
The second project in this thesis aims to augment microrover platforms that are lacking the instruments for distance measurements. Particularly, this project explores the application of monocular depth estimation where the goal is to estimate a depth map from a monocular image. This problem is inherently difficult to solve given that recovering depth from a 2D image is a mathematically ill-posed problem, compounded by the fact that the lunar environment is a dull, colourless landscape. To solve his problem, a dataset of images and their corresponding ground truth depth maps have been taken at Mission Control Space Service's Indoor Analogue Terrain. An autoencoder was then trained to take in the image and output an estimated depth map. The results of this project show that the model is not reliable at gauging the distances of slopes and objects near the horizon. However, the generated depth maps are reliable in the short to mid range, where the distances are most relevant for remote rover operations.