In this project, multiple heuristic-based and data-driven active vision strategies are presented for viewpoint optimization of an arm-mounted depth camera to aid robotic grasping. These strategies aim to efficiently collect data to boost the performance of an underlying grasp synthesis algorithm. An open-source benchmarking platform was created in simulation ( https://github.com/galenbr/2021ActiveVision), and an extensive study for assessing the performance of the proposed methods as well as comparing them against various baseline strategies was performed. The experimental study was done on a Franka Emika Panda robot with a two-fingered parallel jaw gripper. In these experiments, an existing grasp planning benchmark in the literature is utilized. With these analyses, we were able to quantitatively demonstrate the versatility of heuristic methods that prioritize certain types of exploration, and qualitatively show their robustness to both novel objects and the transition from simulation to the real world. The heuristic-based and data-driven methods were also compared in terms of their execution times and efficiency. We identified scenarios in which our methods did not perform well and present a discussion on which avenues for future research show promise.

Creator

• Natarajan, Sabhari

Contributors

• Xiao, Jing
• Gennert, Michael A.
• Calli, Berk

Degree

• MS

Unit

• Robotics Engineering

Publisher

• Worcester Polytechnic Institute

Identifier

• etd-26976

Keyword

• grasp synthesis
• self-supervised learning
• active vision
• benchmarking
• reinforcement learning

Advisor

• Calli, Berk

Committee

• Gennert, Michael A.
• Xiao, Jing

Defense date

• 2021-08-04

Year

• 2021

Date created

• 2021-08-09

Resource type

• Thesis

Rights statement

• In Copyright