Proc. of IEEE/ASME Int. Conf. on Advanced Intelligent Mecharionics, Kobe, Japan, July 20-24, 2003 1 Dav Developmental Humanoid: The Control Architecture and Body Shuqing Zeng, David M. Cherba, Juyang Weng Department of Computer Science and Engineering Michigan State University East Lansing, MI 48824, USA zengshuq, cherbada, weng @cse.msu.edu August 31, 2003 Abstract Dav is a developmental humanoid built in the Embod- ied Intelligence Laboratory at Michigan State University. The framework of Dav’s control architecture is designed by researchers but the actual controller is developed au- tonomously (learned without a programmed task-specific representation) from its own experience. Dav’s percep- tion driven developmental control architecture is differ- ent from a conventional robot controller in following ma- jor aspects: (1) No task (goal) is given at programming time. (2) The task-specific representation is generated au- tonomously from sensor-motor space, not in a 3-D world space. (3) The robot learns while performing. (4) The robot is “alive” while humans interact with it. A devel- opmental robot raises several challenging requirements for its body design that are not the same as those for tradi- tional robots. We outline the major issues in designing the body of such a developmental robot, e.g., mobility, unteth- eredness, sensor and effector requirements, computational resources, longevity, and body wiring. 1 Introduction Recently technological advances have enabled hu- manoid robots to be studied actively in the research com- munity. The first humanoid robot is “Elektro the Motor- man” shown at the 1939 New York World Fair. With the ar- rival of computers, more impressive humanoid robots have been constructed, such as Wabot-2, WABIAN, Cog, Robo- naut and P2. More recent humanoid systems include H6, ETL-Humanoid, ASIMO and Sony SDR-4X. Most of the projects primarily focused on the challenges in the design of mechatronic system and its control as well as integration of anthropomorphic components for a humanoid body. Traditional robots rely on human supplied equations of dynamics and kinematics to plan trajectories. They re- quire a human supplied representation of the environment. This approach requires reliable, high precision body com- ponents with known dynamics and kinematics. Robots designed with this approach have difficulty in adapting to unknown and changing environments, for example, in sensing and controlling a robot’s arm to pick apples in an orchard. Further, the goal can change at any time (e.g., a larger apple becomes the target before the cur- rent one is picked). Some efforts have been made to pro- gram behavior-based humanoid robots (e.g., Kismet [1], Cog [2]). Recently, progress has been made in realizing robots that can develop their mental skills autonomously through what is called Autonomous Mental Development (AMD) [3]. Shown in Figure 1, the AMD paradigm has two phases. In the first phase, the construction and programming phase, tasks that the robot will end up learning are unknown to the programmer. A task-nonspecific program called the developmental program is written in this stage. The sec- ond phase, autonomous development phase, starts when the robot is turned on at time , the robot starts to in- teract with the physical environment in real-time through continuously sensing and acting. Time Given to Agent Ecological conditions Release Turn on ... ... Construction & programming phase Autonomous development phase Sense, act, sense, act, sense, act ... Task 1 Given to Given to Given to Task 2 Task n Training, testing Training, testing Training, testing Figure 1: Autonomous development paradigm. This article describes the design of control architecture and the body of the Dav developmental robot, which is the next generation platform of the developmental robot after SAIL [4]. We outline a developmental perception driven control