Multisolution Bayesian optimization for robotic manipulation tasks.

Herrera Seara, Ignacio
Martínez Cantín, Rubén (dir.)

Universidad de Zaragoza, EINA, 2023
Departamento de Informática e Ingeniería de Sistemas, Área de Ingeniería de Sistemas y Automática

Máster Universitario en Robótica, Gráficos y Visión por Computador

Resumen: Humans have acquired over thousands of years an incredible ability to grasp and manipulate objects, both known and unknown, of different shapes, sizes and materials. This capability to interact with objects is of great interest in robotics. The use of robotic arms with hand-type grippers allows the automation of tasks previously performed by humans. This gives them greater flexibility when manipulating and grasping objects compared to traditional grippers used in the industry. However, replicating this dexterity is not a simple task and, despite recent advances, it is still far from matching human level. One of the main challenges facing this field of research is the grasping of previously unknown objects, which in practice leads to a high overhead in costs due to the required reconfiguration and reprogramming the optimal grasp for each new object. Determining the optimal grasp through a brute force approach is unfeasible due to the cost of operating the real robot. Furthermore, despite the existence of simulators that allow to efficiently evaluate and verify grasps, automating the identification of the optimum remains a complex task. On the other hand, depending on the task to be performed or the existence of occlusions in the environment, it is possible that the optimal grasp may prevent certain manipulations with the object. Therefore, it is necessary to find multiple alternative grasps, sufficiently good and diverse, that allow the correct manipulation of the object. In this work we focus on the problem of automatically obtaining multiple grasps for previously unknown objects with robotic hands. To obtain the grasps, only tactile exploration of the object was used, taking advantage of the sensors in the robotic hands. For this purpose, in this work we have designed and implemented a 3D grasp evaluation environment in simulation with Simox simulator. The iCub robot, designed for research tasks, has been used to carry out experiments. Finally, obtaining multiple grasps has been approached as an optimization problem, where the function to be optimized is a quality metric of a given relative hand-object position. To address this problem we have used Bayesian optimization, an efficient global optimization method that does not require prior information about the problem and works sequentially with a trial-and-error approach to identify the optimum. However, this algorithm is focused on finding a single optimal solution, so it has been necessary to design and implement a variant that identifies multiple solutions (optimal or suboptimal), known as multisolution Bayesian optimization.

Tipo de Trabajo Académico: Trabajo Fin de Master