Decision Making Software for Nuclear Facility Decontamination and Dismantlement, page 2

Decision Making - While most roboticists will intuitively agree that multiple criteria assess the true performance of a robot, incorporating these criteria in real-time decision making is a challenging pursuit. This report describes a method that combines closed-fonn reverse position analysis, nullspace exploration, and a multicriteria evaluation process. The closed-form reverse position analysis satisfies the placement constraints on the robot's End-Effector (EEF). The process of null-space exploration generates configuration options. The evaluation process uses the performance criteria to rank the options and choose one as the recommended configuration. Many researchers have studied this decision making problem in the context of resolved motion control for redundant robots. The coupled operational mode described in this section is an example of resolved motion control. Most resolved motion control work derives from Whitney's resolved motion rate control that suggests the use of the pseudo-inverse for redundant robots. Liegeois showed the extension of this method to include self-motions via the null-space. Since then, a large number of researchers have implemented pseudo-inverse based methods and others have studied their limitations. Other rate control approaches include: Seraji's configuration control, Baillieul's extended Jabobian, the Jacobian transpose and a number of numerical optimization methods such as the series of unconstrained minimizations technique. Dubey and Luh include task-based performance measures in the redundancy resolution. This decision making work applies an entirely different redundancy resolution method. By incorporating closed-form reverse position analysis, this method leverages two decades of work since Freudenstein's declaration of general six DOF reverse position analysis as the Mount Everest of kinematics problems. Other researchers have also shown the use of closed-form reverse position analysis to solve 6 DOF substructures within a redundant robot, though they leave the decision making to a human operator.

The decision making algorithms fall into two general categories depending on the operational mode. In the configuration control mode, the interface suggests one configuration based on task requirements and information on the environment. The software generates the suggestion using a simulated annealing algorithm. In the coupled mode, the software continuously updates the set of joint displacements as the operator "steers" the robot's hands with a manual controller. For this mode, the software employs a sequential filters optimization algorithm. Implementing the configuration advisor presents a challenging decision making problem. Given the desired toolpoint locations, the varied set of performance criteria, and a complex environment with multiple obstacles, the advisor must identify a single solution within the DAWM's 17 degree of freedom solution space. This is a geometrically complex global optimization problem in an extremely large solution space. There are a number of options for finding these global optima including: a “shotgun” approach tracking gradients from different starting places in the workspace, simulated annealing based on a model of the physical annealing process, genetic algorithms based on a model of biological genetics, brute force exhaustive evaluation, and the Monte Carlo based on randomness and statistics. All of these methods will solve global optimization problems. The difficulty lies in the need for interactive response (a few seconds) from the configuration advisor. This section discusses results for the simulated annealing method. The results show this to be a reliable approach, even in complex environments with multiple obstacles and competing performance criteria. Annealing describes a process of heating a material to an elevated temperature and then cooling it very slowly. The slow cooling allows the material to reach a low energy state in which it is relatively ductile. With no intelligence or systematic strategy, some materials minimize their energy state during slow cooling.

Simulated annealing is an approximation of this natural process carried out on a computer and is based on the Boltzmann probability distribution. Boltzmann's constant, and T is the temperature. Essentially, the Boltzmann probability distribution states that a system's energy is probabilistically distributed depending upon the temperature. As the temperature increases, the probability of the system assuming a higher energy state increases. As the temperature is lowered, the odds of the system leaving a lower energy state decrease. Because simulated annealing algorithms sometimes leave lower energy states for higher ones, they can escape from local minima. Simulated annealing algorithms typically include a method of generating random changes in the system's configuration. As applied to the configuration control problem, these changes represent trial configurations options for the robot. The algorithm evaluates these options using the Boltzmann probability distribution. If the distribution indicates, the system assumes the trial configuration; otherwise it is discarded. The coupled motion mode is the mode most often developed in the robotics literature on redundancy resolution. This section discusses a method of redundancy resolution that uses local exploration to explicitly identifv a set of options for the robot's motion and then applies sequential filters to identify one option as a local optimum. Simulated perturbations at the joint level drive the exploration. Eschenbach and Tesar developed the sequential filters method and applied it to the mechanism synthesis problem. They reported an example of the method reducing a design space of 60,000 to one of only 50 ranked solutions. A similar approach using a joint-level hypercube can generate configuration options for redundant robots. There are 2n points on the faces of the cube, 2 to the n points at the vertices, and 3 to the n points in all. Respectively, these are the simple, factorial, and exhaustive exploration patterns. After exploration generates the set of options, sequential filters evaluate and rank the options based on the performance criteria and operational constraints. The logical sequence first applies the least computationally demanding constraint criteria, and then evaluates the remaining options based on the higher-level performance criteria. This technique will resolve the redundancy of extremely complex systems at hundreds of cycles per second on common personal computers.  Next Page ->