Paper 155 Details

Title

Shaping the body to shape the behavior: a more active role of the morphology in the brain-body trade-off

Abstract

In recent years the concepts of embodiment and morphological computation brought an important paradigm shift in a number of research fields such as robotics, artificial intelligence, and artificial life. One of the most remarkable implications of these theoretical frameworks is the idea that the body has indeed an active role in the emergence of intelligent and adaptive behavior. It has been shown that in many cases it is possible to achieve complex behaviors with no ”brain” at all (i.e. with no control – as in the case of passive dynamic walkers, or with a simple, periodic, open-loop control – as in the case of many walking and running robots). This view opens a number of intriguing questions, like: to what extent can we outsource intelligence to the body? What kind of behaviors can we achieve without introducing control? These questions fall under what is sometimes called the brain-body trade-off. With this abstract we aim at contributing to such a discussion by presenting some recent results of our work from an original perspective. The ultimate goal is to show that the body can have a more active role in achieving diverse behaviors. To stress this concept, we operate at one of the extremes of the brain-body continuum in which the role of the brain (or controller) is close to zero, i.e. the one of self-stabilizing robots with periodic open-loop control. There are many examples of such robots in robotics research, characterized by some common features. First, those robots have a morphology that is typically carefully engineered by human experts, often following heuristic design criteria and trial and error tests. This appears to be limiting if we consider the complexity of a robot’s design space and the fact that such approaches are not able to explore many possibilities with respect to morphology and behavior. Second, although facilitating control, morphology usually has a passive role: once the design is fabricated morphology stays unchanged over the life span of the robot, and often operates in one or few, fixed, working regimes (or attractors, in a dynamical systems’ view). Third, their behavioral diversity is usually limited, and where it is not the case, it is achieved by varying the controller, not by exploiting the morphology. Here we propose a systematic procedure that allows to design self-stabilizing robots exploiting the full potential of the morphology to achieve a diversity of behaviors. The key idea is to actively and dynamically involve the morphology in producing diverse behavior, resorting to the concept of morphosis or morphing, i.e. the possibility for a robot to control/experience a morphological change during its life span. The idea is to allow the robot to change its shape to change its behavior, in presence of a fixed controller: a walking robot may morph to switch to a running gait, while an arm performing a limit cycle movement (e.g. exploration behavior) may morph to perform a reaching or a grasping movement. Given a basic morphological structure, an evolutionary process maximizing a metric of behavioral novelty is first ran to explore the space of morphology-enabled behaviors. This process allows to simulate a very large space of robot configurations without the restriction and the biases of the common trial and error, heuristic design procedures. Then an automated clustering procedure is executed to group the results of the evolutionary process into subsets of similar morphologies. Inside these clusters we search for functional robot configurations that are close in the morphology space, but far apart in the behavior space. Those configurations are selected as candidate morphologies to apply morphing, implemented as a gradual transition from one morphology to another: this way a slight morphological change can result in a macroscopic change in the behavior. In what follows some details regarding the procedure and the achieved results are provided. It is worth noting that there are considerable differences among this study and others in which novelty-based search is adopted to evolve morphologies (the closest one being Lehman and Stanley (2011b), where the focus is on combining the pressure towards novelty with the one towards performances). In addition to differences in the goals and in the applied methodology, the most notable, general, difference is that these works do not consider morphing, being instead the core of this work. For further details the interested reader is encouraged to refer to Corucci et al. (2015).