Title of Presentation: Title: Self-Supervised Terrain Classification for Planetary Rovers
Primary (Corresponding) Author: Christopher Brooks
Organization of Primary Author: Massachusetts Institute of Technology
Co-Authors: Karl Iagnemma
Abstract: For planetary rovers, autonomous mobility is a key to enabling greater science return, through both an increase in the range of environments safely traversable and a decrease in the number of command cycles to get there. Mobility in rough terrain can be greatly affected by terrain properties, especially near craters, cliffs, and ravines, all of which are targets of scientific interest. For example, while a rover may climb a rocky slope with ease, it may slide down a sandy slope of the same grade. Even on flat ground, a rover may become entrenched in soft soil. The ability to remotely sense these terrain properties could allow a rover to avoid potentially-hazardous terrains, enabling more autonomous navigation while reducing the potential for loss of mobility. In previously-unexplored environments, however, the appearance of terrain classes may not be known a priori. In these situations, a rover should learn from its past experiences with different types of terrain to predict where it may safely travel in the future.
This work describes a self-supervised classification approach to learning the visual appearance of terrain classes. Here, the local terrain class is identified during normal driving based on vibrations induced in the rover structure, using a previously-trained vibration-based terrain classifier. By combining the local terrain identification and the recalled appearance of the local terrain, a vision-based classifier is developed to recognize the terrain classes. This vision-based classifier can then be used to classify terrain in the full field of view of the rover's cameras.
Experiments were performed using a four-wheeled rover in Mars analog terrain. Results demonstrate the potential for this approach to improve the knowledge of the properties of the surrounding terrains.