Metareasoning for Robots: Adapting in Dynamic and Uncertain Environments

Jeffrey W. Herrmann

Professor, University of Maryland, College Park

Metareasoning for Robots: Adapting in Dynamic and Uncertain Environments (Springer, 2023) is a state-of-the-art resource that robotics researchers and engineers can use to make their robots and autonomous vehicles smarter. Readers will be able to describe metareasoning, select an appropriate metareasoning approach, and synthesize metareasoning policies.

For ordering information, see the publisher web site.

Metareasoning for Robots: Adapting in Dynamic and Uncertain Environments contains the following chapters:

1. Introduction to Metareasoning: This chapter introduces some key concepts related to robots and autonomous systems. It then discusses why metareasoning is needed and the benefits of metareasoning, which is a branch of artificial intelligence (AI). It presents a list of key sources that one should read for more information about metareasoning. Finally, it discusses the systems engineering approach that informs the structure and contents of this book.
2. Metareasoning Design Options: This chapter discusses the options that are available for metareasoning ap-proaches and policies, using examples from our own work and the literature to il-lustrate the techniques, and the advantages and drawbacks of different options. By making these decisions about the metareasoning approach, we are beginning to specify how the robot will perform metareasoning. Naturally, we might find that these choices do not work as expected, and we might need to iterate and revisit these decisions. This chapter describes many ways that a robot can use metarea-soning to improve performance, reliability, and safety.
3. Implementing Metareasoning: This chapter describes options for locating the software that will implement the metareasoning process on a robot. The details of the implementation depend upon the robot's autonomy software architecture, so this chapter reviews some common architectures. This chapter also discusses three possible locations for metareasoning: (1) as a separate software that runs in parallel to the autonomy software, (2) as a separate module that is added to the autonomy software, and (3) as additional code that is added to one or more parts of the autonomy soft-ware. This chapter also discusses the details of a specific implementation in the autonomy stack of a mobile ground robot.
4. Synthesizing Metareasoning: After implementing a metareasoning approach, the next step is to determine the specifics of the metareasoning policy. The metareasoning policy controls the object level reasoning (by selecting an algorithm or adjusting a parameter value, for instance). This chapter describes a systematic, data-driven approach for deter-mining the metareasoning policy and some case studies to illustrate the process. Although machine learning techniques have been employed to determine metareasoning policies, using engineering judgment, based on empirical results, is an appropriate technique as well.
5. Testing Metareasoning Policies: This chapter discusses ways to measure the performance of a metareasoning policy and approaches for designing the tests that are needed to collect evidence about a metareasoning policy's performance. Assurance cases can be useful to supporting claims about a metareasoning policy's performance and safety. In addition, we should document our testing in a way that supports replicability so that others can repeat our tests if needed. This chapter also describes techniques for visualizing metareasoning, which can be useful for understanding and improving the metareasoning approach.

This page last updated on May 24, 2023.