Hybrid control for autonomous systems - Integrating learning, deliberation and reactive control

High-level control of Autonomous Systems (e.g., robots) is concerned with selecting the next action the system should perform. In particular, this means that the system must be endowed with the capability to derive and execute the next suitable step towards itsmission goal. This problem is a fundamental one in Robotics and Artificial Intelligence, and has existed for as long as the research areas themselves. The well-known classical paradigms for tackling the action selection problem are learning, deliberation, and reactive control. Due to the advantages and drawbacks of the basic approaches, combinations of these schemes are often used. This leads to the research topic of hybrid control, where more than one of the basic paradigms are seamlessly integrated. (The term ‘‘hybrid control’’ is also used in engineering, where it describesmixtures of discrete time and continuous control systems.) Learning has been applied successfully to many robotics tasks. Most of the work is related to learning certain basic behaviours or skills (see e.g. [1]). The advantage of this paradigm is that even complex behaviours do not need to be explicitly programmed, but can be taught from scratch, or optimised with repeated training. Examples in which the high-level control strategy of robots (or any agents) were successfully learned are scarce. The deliberative approach for decision making of autonomous systems has been successfully treated in research on Artificial Intelligence (AI) for decades, following a top-down approach. The advantage of this paradigm is the intuitive and elegant description of complex tasks and domains [2]. However, this has severe limitations in real applications. For instance, these approaches might lack reactiveness, which is required for the application domain. In the reactive control paradigm the idea is that, by combining purely reactive action selection schemes, intelligent and goal-directed behaviours emerge [3]. This fast and reactive paradigm can be seen as a bottom-up approach. These different paradigms have been known and investigated for several decades, and in fact, in today’s applications, mostly combinations of learning, deliberation and reactive control are used. Usually these combinations are used in an ad hoc or even unconscious fashion. Although there are a number of proposed architectures and a huge body of literature, the issue of combining learning, reactive control, and deliberative control has never been intensively and systematically investigated. In order to initiate a deeper discussion on this topic and to establish a more sound foundation, we organised the 1st workshop on Hybrid Control of Autonomous Systems — Integrating