Task dependent dexterous hand control

摘要

We will probe deep into the issue of task dependent dexterous hand control. We will explain the meaning of the word task, bring forth the dependence of hand-object-environment dynamic models on tasks, and develop task dependent control strategies for their execution on a chosen hardware configuration. The hand used is the JPL/Stanford Hand, run on a single MicroVax using the VaxELN operating system, through interface hardware built inhouse. We begin this thesis with an analysis of mechanical contacts and develop formal definitions and representations for a set of six task primitives for general manipulator usage-free, guarded and fine motion, and force application. We explain how this vocabulary is useful in describing interactions between a single finger of a dexterous hand and its environment. We then construct five hand task primitives--preshape, grasping, acquisition, motion and force application manipulation--from the six finger task primitives and two kinds of inter-finger dependencies--synchronocity and coordination between fingers. Task primitive dependent dynamics models are developed using the theory of Dynamics of interconnected systems. Task dependence is brought out through a characterisation process. As a result, the hand (robot) and the object (environment) may require to be modelled as in the lagrangian form or the co-langrangian form. Note that all such models must be in the manipulator's operational space. In the actual development of dynamic models, we recognize that most mechanical systems are inertial systems, whose co-lagrangians are extremely difficult to develop from first principles and propose a pseudo form, far easier to develop. An algorithm is presented for its development and dynamic models for general manipulator usage as well as for dexterous hand usage. The next part deals with task dependent control strategies for dexterous hands. Control architectures are developed for preshaping, grasping, acquisition and manipulation actions. The whole theory is illustrated through the usage of the JPL/Stanford Hand in turning a bolt. Pole placement methods using the multivariable feedback control techniques are used for compensator design. Some preliminary implementation results are presented. The advantages of the approach here is discussed and some future directions pointed out. (Abstract shortened with permission of author.)