Epsilon Optimal Path Planning for Active Vision for Grasping

Brown, Galen

Etd

Epsilon Optimal Path Planning for Active Vision for Grasping

Public Deposited

In this work I explore the use of active vision algorithms to improve robotic grasping. Robotic grasping algorithms aim to find a suitable grasp location on a given target object using visual data, i.e. images taken from a camera. Their performance depends on the camera viewpoint; some viewpoints are more suitable to detect a good grasp location than others. The role of active vision is to alter the camera viewpoint to help the robotic grasping algorithms. I provide an extensive overview of active vision strategies to find ‘sufficiently good’ grasps with as little camera movement as possible. I present several heuristic and data driven approaches to the problem in a constrained, discretized scenario and compare their performance. In addition, I demonstrate a method to find solutions in a more realistic, continuous scenario that are within a known error bound of optimal solutions. I outline the mathematical basis for this claim, and demonstrate its empirical characteristics in a number of simulated experiments. Using this information, I am able to show the limitations of current approaches and demonstrate that significant improvements in performance can be made by working in the continuous space rather than constraining the problem to the discrete space. This work provides novel information about the theoretical limits of active vision, which suggest directions for future research.

Creator