Autonomy in automation: trends, technologies, tools

Any but the most indomitable technophile today would, we expect, refuse to be a passenger in a pilotless airliner or have an uninhabited yet operational refinery in her neighborhood. But, regardless, the trend toward reducing human involvement in the operation of complex engineering systems, driven as it is by considerations of economics, performance, and safety, appears inexorable. Further substantial improvement in process automation, however, will require more than evolutionary technological advances. Our focus in this paper has been on autonomy, a property absent in today's automation systems. In order to satisfy the demands of industry and society, we will need to make our automation solutions autonomous-they will need to be able to respond appropriately to unforeseen situations, not just limited to precompiled behaviors. We have noted two research directions that are central for engineering autonomy. First, diverse representations of knowledge-of process dynamics, failure modes, the environment, and other factors-will be needed and will need to be integrated so that unanticipated situations can be explicitly reasoned about. Models are widely recognized as principal determiners of control system performance; we must broaden our attention now to multimodels. Second, dynamic resource management technology must be developed to allow multiple, competing, heterogeneous computational tasks to execute on real-time platforms under hard and soft deadlines and resource constraints. The full space of relative priorities that could be faced by an autonomous system cannot be predicted; tradeoffs among resource allocations for different tasks will need to be made online and adaptively, not through static scheduling. Engineering autonomy will require maturing concepts into tools. Tools that can truly enable multimodeling and dynamic resource management are unavailable today. A key, general differentiating attribute that distinguishes the tools that are needed from those that are currently in use is the level of integration. Whether it is models of different subsystems, or knowledge from disparate sources, or performance and resource usage aspects of computational tasks, tools for autonomous systems will require integration of diverse phenomena, features, and representations. Finally, we suspect that our pursuit of autonomy in automation and control will bring the topic of consciousness-at least in the limited sense of an agent's awareness of its environment and context and its deliberative adaptation of its information processing priorities-into the foreground of discussion.