Poster Session 103 Minimalistic Behavioral Rule for Reflecting Robot’s Morphology Fabio Dalla Libera 1 ,Shuhei Ikemoto 2 , Koh Hosoda 2 and Hiroshi Ishiguro 3 1 Research Fellow of the Japan Society for the Promotion of Science (E-mail: fabiodl@gmail.com) 2 Graduate School of Information Science and Technology, Osaka University, Osaka, Japan (Tel: +81-06-6879-7739; E-mail: {ikemoto, koh.hosoda}@ist.osaka-u.ac.jp) 3 Department of Systems Innovation, Graduate School of Engineering Science, Osaka University, Osaka, Japan (Tel: +81-06-6850-6360; E-mail: ishiguro@sys.es.osaka-u.ac.jp) Abstract: In a previos work, we proposed a very simple stochastic model, termed Minimalistic Behavioral Rule, in order to show how small bacteria such as Escherichia coli can robustly reach high concentrations of nutrient despite the noise in the sensory information. In particular, we showed that when this simple behavioral rule is employed, environmental or internally generated noise can be beneficial to the resultant behaviors of the living being, a phenomenon that can be explained by Stochastic Resonance. In this paper, we apply such behavioral rule to a real world complex robot, whose behavior is strongly influenced by its morphology and its surroning environment. Through the experiments, in particular, we show that the sensory information used for the task achievement greatly influences the resultant behavior. Keywords: Minimalistic Behavioral Rule, Musculoskeletal robot arm, Adaptive behavior 1. INTRODUCTION Living things can survive in complex and dynami- cal environments by taking full advantage of their body dynamics, sensing and interaction with the surrounding environments. Small bacteria such as Escherichia coli (E. coli) are no exception. In a previous work, we pro- posed a Minimalistic Behavioral Rule (MBR) in order to explain how E. coli can effectively reach high con- centrations of nutrients and avoid high concentrations of repellent substance despite highly noisy sensory infor- mation[1]. Since MBR is extremely simple and makes very limited assumptions, it can be easily applied with- out knowing the robot’s body structure or its actuators properties. Experiments showed that MBR can control simple mobile robots with no information on its actuators and sensors [2]. However, to date, MBR was not tested on complex, multi-DOFs robots. The idea of applying a very simple control to highly complex robots is not new. So far, many researchers have developed biologically inspired robots [3] that can oper- ate with simple control laws. Usually, the exploitation of the morphological computation[4], emergent from a well-designed robot’s body, allows the achievement of a specific task with very simple control laws. However, the identification of such simple control laws requires the developer’s inspiration, knowledge and experience. In other words, even if the control laws are very simple, it is not easy to find them. The ultimate goal of this research is to build a simple but general control law which can exploit the character- istics of the robot’s morphology automatically. We pro- pose MBR as a possible solution for controlling a robot when no previous knowledge on the robot’s actuators and sensory data is available. If specific knowledge is avail- able, clearly, task and robot specific controllers can be de- signed to improve the system efficiency. Actually, MBR can be used for collecting the data necessary to this de- velopment process. Fig. 1 The complex musculoskeletal robot arm used in the experiment. In this paper, we show that MBR is applicable even when the robot has a very complex structure. In detail, MBR was used to control the pneumatic musculoskele- tal robot arm shown in Fig. 1. This robot has a 7 DOFs driven by 17 McKibben pneumatic muscles. Each mus- cle is equipped with a pressure sensor, used for closed loop pressure control. In other terms, the robot is con- trolled by setting the variation of the pressure in each of the 17 pneumatic actuators. The task chosen consists in reaching three points in sequence. In the experiments, the sensory information available to the robot was changed in the experiments, to observe differences in the behavior. 2. MINIMALISTIC BEHAVIORAL RULE In [2] we proposed the Minimalistic Behavioral Rule: u i t+1 =  u i t + η i R if ΔA t ≥ 0 random selection otherwise . (1) Where the u i t indicates the i-th component of an m- dimensional motor command u i t ∈ R ⋗ given at time t