Approaching human performance: The functionality driven Awiwi robot hand

Kurzfassung

Humanoid robotics have achieved a remarkable state in recent years. Nowadays humanoids can walk stairs, serve coffee, throw and catch balls and interact with human beings. However, most of these demonstrations and applications take place in well known environments or even in surroundings that have been adapted to the robots capabilities and needs. However, in order to assist the human in every day tasks, the robot has to operate in (partially) unknown environments in most cases. In these unknown environments and in interaction with moving obstacles as well as human beings, collision avoidance is vague notion. Consequently, this dissertation hypothesizes that the operation of humanoid robots outside of environments dedicated to operate the robots implies that robots have to be able to complete tasks even in case of collision. This especially applies to robot hands, since they are the most exposed and fragile part of a humanoid robot. Humanoid robots have to be anthropomorphic in sense of providing not only human-like appearance but also human characteristics. In particular they have to provide: Robustness against impacts, fast dynamics, human-like grasping and manipulation performance To achieve this robustness and fast dynamics, from the author's point of view, a paradigm change has to be done . Future robots have to be able to store energy as suggested by T. Morita [Morita et al. 1999]. In this thesis the anthropomorphic Awiwi Hand is developed, which provides human-like robustness and dynamics as well as grasping performance. To achieve these characteristics, the human anatomy as well as existing robot hands are analyzed. The goal of this analysis is to derive the functionalities needed to achieve real anthropomorphism rather than to blindly copy the human being. These abstract functionalities are then implemented to a robotic hand. The achieved anthropomorphic characteristics of the Awiwi Hand are demonstrated in several experiments. The Awiwi Hand is able to withstand the impact of a 500 g hammer at high velocity without any damage. It can still keep objects firmly grasped even when struck by an 750 g object at a speed of approximately 4 m/s. The energy stored in the elastic elements of its antagonistic drive train allows the fingers of the hand to achieve a maximum finger speed of approximately 3500 grad/s which is more than five times the speed provided by the drives alone. The Awiwi Hand is, to the author’s knowledge, the first robot hand able to perform all grasps of M. Cutkoskys grasp taxonomy [Cutkosky 1989]. The robustness, fast dynamics and grasping performance of the Awiwi Hand is thought to enable future humanoid robots to operate in "field robotics" rather than in laboratories built for the robots. It will speed up the development of robotic applications since developers will no longer have to bother to avoid possibly costly collisions of the robot. Methods such as reinforcement learning, which need failed task execution attempts to succeed, can be used without fears of severely damaging the robot. The method underlying this development is not limited to robot hands. The proposed methodology will help realize a new generation of humanoid robots that can assist the human being even in harsh environments without damage and for example might fall over without damage. They will hopefully accommodate the demand of the human society for robot assistants that is well documented by the public interest in humanoid robotics.