A Machine Learning-Based Approach for Traversability Analysis and Path Planning in Martian Rovers

Artificial intelligence is considered one of the major breakthroughs in data analysis and decision-making of the last decades. It is one of the scientific disciplines that have most attracted the attention of the public and the scientific community, with applications ranging from the medical field to cybersecurity. With the advent of Internet-of-Things and the increasing number of high-quality sensors present inside most of market-dominating smartphones, the need for a class of methods capable of extracting fundamental information from an unprecedented volume of data was, and still is, very high. Machine and Deep Learning methods are the main instruments used nowadays to perform such intricate analyses, allowing to establish relationships between input features and output predictions in a similar way to the human brain, without requiring extensive training of human operators to repeat the same analysis on new data. Interplanetary exploration through autonomous rover systems is a perfect example of a scientifically challenging task that can greatly benefit from the application of Machine Learning algorithms within its workflow. While new solutions are being devised in order to allow humans to travel, land and conduct scientific research on the Martian soil, Martian rovers represent one of the essential elements that enable the exploration of the red planet surface. One of the main factors of improvement for the Martian rover missions is the velocity of exploration, i.e. to move faster while avoiding obstacles and unstable terrains. To meet this objective, drones and satellites can provide significant information for the rover navigation in terms of maps definition and planning of the path. This objective poses many technical challenges such as autonomous terrain labeling, traversability analysis and path planning. Solutions for the two latter tasks have been studied and implemented in this thesis. Traversability analysis is the quest of examining the data related to a portion of terrain in order to establish which parts of it are safe for the rover to traverse, considering different constraints regarding terrain composition, slope percentage and hazard presence, in order to output a traversability map that can be fed to the planning subsystem. Path planning, instead, is the task of searching for the optimal sequence of valid configurations that makes the system reach the goal point, given its current location, a representation of the environment that includes the hazards (i.e. the result of traversability mapping) and the heuristic that we want to use to establish what it means for a set of configurations to be optimal. Both of these tasks have been performed with a learning-based approach, as the main assumption behind this work is the fact that machine learning algorithms can be very helpful to carry out such analyses, that require using vast amounts of data that are difficult and time consuming for humans to manipulate. The research regarding traversability mapping was carried out within the SINAV project in collaboration with Altec S.p.A. and sponsored by the Italian Space Agency, under the supervision of Chiara Leuzzi, while the path planning part was conducted within the research environment of the Politecnico di Torino and supervised by Professor Fabrizio Stesina.