Structuring learning from demonstration to support verifiable robot control

With the increasing level of adoption of autonomous systems in our daily lives, as robots make their way out of the restricted factory floors into our hospitals and shopping malls, there is now a need to consider how they should be designed so that they may be trusted. Achieving trustworthiness involves several complex criteria, but a useful description of the desiderata in a form that is compatible with the engineering design process is to ask whether a system will function as we say it should, whether we know how it will fail and if we have evidence regarding how it continues to function after it has failed to satisfy some criteria. Learning and adaptation are central features of many modern autonomous systems. In particular, the paradigm of Learning from Demonstrations is a widely used approach in numerous systems being deployed in fielded applications. Establishing whether these desiderata are satisfied by a controlled autonomous system based on learning from demonstrations can be particularly difficult. We argue that this is best achieved when models and representations used by the learning algorithms are suitably structured and when the learning paradigms explicitly incorporate input from human experts or end users. In this chapter, we outline an approach to achieving this, and we describe some techniques that aid in such a design process. This is grounded in examples taken from the domain of assistive operations with a humanoid robot, but similar principles apply also in most other applications of robotics and autonomous systems.