Control Systems and the Quest for Autonomy

Autonomous vehicles have certainly captured the imaginations of everyone. The promise of reducing or even eliminating accidents via autonomy is very appealing. The quest for autonomy has been a pervasive theme in engineered systems throughout the centuries. Adding advanced sensing and decision making to traditional control systems is a way to increase the level of autonomy of a system. Control systems are seen as a cornerstone of autonomous dynamic systems. A functional hierarchical architecture describing the necessary functions needed in an autonomous space vehicle is presented. These concepts were developed while the author was visiting research faculty at NASA’s JPL.

When people refer to autonomous systems, they often mean different things. It is important to be more precise and agree upon a common definition: Autonomy is the ability of a system to achieve a set of goals under uncertainty in the system and its environment. For the same set of goals, the larger the uncertainties the system can handle, the higher the degree of autonomy. The lower the needed external intervention by humans or other systems to achieve the goals under the uncertainties, the higher the degree of autonomy. So, the level of autonomy depends on both the measure of the set of the goals that are being accomplished and the measure of the set of uncertainty present. Specifically, {measure of the set of goals} x {measure of the set of uncertainties} = L, the level of autonomy. This definition allows the comparison of the autonomy levels of different systems.

Note that since every autonomous system has a set of goals to be achieved under a set of uncertainties, and a control mechanism to achieve them, clearly, every autonomous system is a control system.

Adding advanced functionality to controllers is essential to achieving higher levels of autonomy. Learning is very important to autonomy. Online learning control for autonomy should rely on prior knowledge, use simplified models, take full advantage of smart data, and incorporate active learning. The following is a list of directions that need to be emphasized more if online learning control is to become successful in generating correct improved controllers in reasonable amount of time:

• Use simplified models with fewer parameters as needed for goals.
• Use all available prior knowledge for the model and use smart initial conditions.
• Use smart data, loaded with pertinent information, and incorporate active learning to go after what you need.
• Use trusted data; trust but verify.
• Time and resources are limited; do your best under limited resources and stop earlier before reaching the optimum, if needed.