OVERVIEW: 

The use of real-time autonomy in vehicles is hampered due to the underlying issue that all algorithms and software used for assessing situations can suffer from the problems of missed detections and/or false positives – that is, missing some aspect of the true situation that is important, or calculating that some aspect is true about the situation that is, in fact, not true. This issue occurs in all applications of autonomy from autonomous cars navigating through traffic signals to UAV’s flying in clouds, fog, or smoke. 

While improving the sensors that feed information to these systems can certainly improve the resulting behaviors, there is also the possibility that using machine learning approaches can improve the tradeoffs between missed detections and false positives in the software used to assess situations, resulting in a better, more trustworthy system.

For space applications, this issue arises in a number of problems: for example, techniques used for on-board detection and tracking of nearby spacecraft and calculation of their orbital trajectories, to determine the threat of collision; assessment of the vehicle’s own state of health and it’s ability to perform requested tasks in a safe manner without compromising its own safety based on the data is collects on its own subsystems, such as power, temperature, etc.

We are interested in how machine learning approaches might be applied to making on-board, autonomous execution of a variety of spacecraft functions achieve more robust performance than current approaches do, which in turn will increase the trust operators have in this on-board autonomy. Leveraging developments and approaches from autonomous platforms being developed for other domains (e.g., self-driving cars, autonomous UAV’s) to show how such methods can help on-orbit vehicles is potentially a key technology to enable autonomous satellites in the future.