In the Task-Driven Generation of Preventive Sensing Plans for Execution of Robotic Assemblies
Conant Pablos, Santiago E.
MetadataShow full item record
It is well known that success during robotic assemblies depends on the correct execution of the sequence of assembly steps established in a plan. In turn, the correct execution of these steps depend on the conformance to a series of preconditions and postconditions on the states of the assembly elements and in the consistent, repeatable, and precise actions of the assembler (for instance, a robotic arm). Unfortunately, the ubiquitous and inherent real-life uncertainty and variation in the work-cell, in the assembly robot calibration, and in the robot actions, could produce errors and deviations during the execution of the plan. This dissertation investigates several issues related to the use of geometric information about the models of component objects of assemblies and the process of contact formation among such objects for tackling the automatic planning of sensing strategies. The studies and experiments conducted during this research have led to the development of novel methods for enabling robots to detect critical errors and deviations from a nominal assembly plan during its execution. The errors are detected before they cause failure of assembly operations, when the objects that will cause a problem are manipulated. Having control over these objects, commanded adjustment actions are expected to correct the errors. First, a new approach is proposed for determining which assembly tasks require using vision and force feedback data to verify their preconditions and the preconditions of future tasks that would be affected by lack of precision in the execution of those tasks. For this, a method is proposed for systematically assign force compliance skills for monitoring and controlling the execution of tasks that involve contacts between the object manipulated by the robot arm in the task and the objects that conform its direct environmental configuration. Also, a strategy is developed to deduce visual sensing requirements for the manipulated object of the current task and the objects that conform its environment configuration. This strategy includes a geometric reasoning mechanism that propagates alignment constraints in a form of a dependency graph. Such graph codes the complete set of critical alignment constraints, and then expresses the visionand force sensing requirements for the analyzed assembly plan. Recognizing the importance of having a correct environment configuration to succeed in the execution of a task that involve multiple objects, the propagation of critical dependencies allow to anticipate potential problems that could irremediably affect the successful execution of subsequent assembly operations. This propagation scheme represents the heart of this dissertation work because it provides the basis for the rest of the contributions and work. The approach was extensively tested demonstrating its correct execution in all the test cases. Next, knowing which are the tasks that require preventive sensing operations, a sensor planning approach is proposed to determine an ordering of potential viewpoints to position the camera that will be used to implement the feedback operations. The approach does not consider kinematic constraints in the active camera mechanism. The viewpoints are ordered depending on a measure computed from the intersection of two regions describing the tolerance of tasks to error and the expected uncertainty from iii an object localization tool. A method has been posed to analytically deduce the descriptions of inequalities that implicitly describe a region of tolerated error. Also, an algorithm that implements an empirical method to determine the form and orientation of six-dimensional ellipsoids is proposed to model and quantify the uncertainty of the localization tool. It was experimentally shown that the goodness measure is an adequate criterion for ordering the viewpoints because agrees with the resulting success ratio of real-life task execution after using the visual information to adjust the configuration of the manipulated objects. Furthermore, an active vision mechanism is also developed and tested to perform visual verification tasks. This mechanism allows the camera move around the assembly scene to recollect visual information. The active camera was also used during the experimentation phase. Finally, a method is proposed to construct a complete visual strategy for an assembly plan. This method decides the specific sequence of viewpoints to be used for localizing the objects that were specified by the visual sensing analyzer. The method transforms the problem of deciding a sequence of camera motions into a multi-objective optimization problem that is solved in two phases: a local phase that reduces the original set of potential viewpoints to small sets of viewpoints with the best predicted success probability values of the kinematically feasible viewpoints for the active camera; and a global phase that decides a single viewpoint for each object in a task and then stitch them together to form the visual sensing strategy for the assembly plan.