Resumen de Application and use of quality metrics for the prediction of grasp success and evaluation of artificial hands
- español
La manipulación ha sido una de las grandes áreas de interés en la rootica durante décadas. Encontrar una forma de agarre adecuada para manipular objetos y disear manos capaces de dichos agarres son dos de los principales problemas en este campo. En esta tesis se analia el uso de distintas métricas de calidad para evaluar las hipótesis de agarre y, mediante algoritmos de aprendizaje automico, predecir el éxito de los mismos. Por último, dichos algoritmos son utilizados para evalaur diseños de manos y sus capacidades para agarrar objetos. Los experimentos llevados a cabo en esta tesis se han realizado tanto en entornos de simulación como con robots reales.
- English
Artificial manipulation has been one of the great areas of interest in robotics for decades. Find a proper grasp to seize objects and design robotic hands capable of such grips are two of the main issues in this field.
First, grasp synthesis is focused in generate a proper hand configuration to resolve a manipulation task. This derives in finding a good grasp among the infinite set of candidates. The selection of this configuration implies to use a methodology to evaluate grasps hypotheses. Quality metrics have been widely used in the robotics field to evaluate grasp candidates. However, experiments showed these metrics are not capable to classify properly these grasps as good or bad.
On the other hand, finding a good grasp is also dependent on the gripper capabilities. For this purpose, different robotic grippers and artificial hands have been developed. Although there are notable differences between grippers, there is no proper way to evaluate their performance. Different approaches have been defined to solve this problem, but none of them seemed to give a convincing evaluation.
We consider quality metrics can be applied in this field to provide an evaluation on different gripper capabilities and also, generate a benchmark to evaluate artificial hands.
The work described in this thesis has three objectives: first, study the characteristics of grasp quality metrics. Second, find a combination of metrics capable of predict the success of a grasp hypotheses. And third, the appliance of these measures to evaluate the performance of artificial hands.
In order to achieve these objectives, we perform an exhaustive statistical study on the characteristics of the most common quality metrics. Then, different classification methods are trained to generate a model capable to predict the outcome of real grasp executions. Finally, quality metrics are considered as evaluators of different manipulator properties. A methodology is proposed in order to evaluate different hands and design improvements.
The results of this thesis will provide a better understanding of quality metrics. A model to predict grasp success. And a methodology to evaluate the functionality of artificial hands.
