Editorial: Agency in Natural and Artificial Systems

The 1980s and 1990s saw a reorientation of the field of artificial intelligence (AI) away from disembodied software, a domain where AI research continues predominantly as software engineering, and toward situated and embodied robotics. Crucially, the publication of Braitenberg's (1984) Vehicles and, later on, the behavior-based robotics movement (e.g., Brooks, 1995) catalyzed this emphasis on agency rather than abstract reason, by demonstrating that circuits embedded in closed sensorimotor loops can exhibit adaptive capacities that exceed those of disembodied software. AIBO or Honda's ASIMO), that rely on explicit design and planning, have succeeded in producing lifelike behaviors. Such artifacts appeal to us, attract our interest, and evoke our sympathy. In a 21st century variant of the Turing test (Turing, 1950) that probes our capacity to distinguish an artificial system's embodied agency from natural agency, not its verbal communication skills, such artifacts pass with flying colors. Yet, something still appears to be missing in current robotics. Criticism has been phrased that such machines merely look as if they act, as if they are smart, as if they have intentions (e.g., Di Paolo, 2003) or that Turing-test style approaches that focus on the surface behavior need to be complemented with the study of behavior generating mechanisms to avoid falling victim to a farce (Rohde & Stewart, 2008). The gist of such criticisms is that, unlike humans or animals, current robots do not really care about the success of their actions. Moreover, the situations in which robot behavior breaks down reveal a lack of even understanding the goal of their actions in the first place. Arguably, a robot that moves but does not understand what its actions should achieve is not really an agent in a strong sense, it only acts as if. Biology shows us that it is possible for a physical system to act according to inherent goals, something we are still missing out in current robotics. The question is whether what is missing is another mechanism to be added to current robot architectures, or whether our modeling approaches have to be rethought in a more fundamental way. How can we build machines that possess stronger forms of agency? The goal of this special issue is to advance on this fundamental problem, identifying new principles for the synthesis of agency. The contributions share in common that they identify complex processes of dynamical self-organization on several levels as possible routes toward obtaining genuine …