Automatic controller generation based on dependency network of multi-modal sensor variables for musculoskeletal robotic arm

Autonomous robots that work in the same environment as humans are preferred to ensure mechanical safety with respect to soft contact with their surroundings and adaptivity to handle various tools and to manage partial malfunctions. To ensure that these requirements for robots are satisfied, this study proposes an approach for obtaining a robot structure and its application to building controller for dynamic motion of a robot. It is assumed that the physical relations between the sensor variables are unknown. On the basis of dependency network construction using mutual information, controllers are generated and tested by finding appropriate causal chains of the sensor variables. The proposed controller generation methods were tested using the control tasks of a musculoskeletal robotic arm. Thus, the proposed controller generation algorithm finds appropriate controllers, and the framework of this generation is robust to the changes in the body of the body. (C) 2019 The Authors. Published by Elsevier B.V.