Open-Vocabulary Part-Based Grasping
Tjeard van Oor · Queensland University of Technology · 2024
This thesis developed an approach to allow robots to grasp objects by desired parts using deep learning models. This approach was tested across 1000 real-world experiments where the robot lifted a variety of objects safely by user-specified parts. The thesis evaluated a variety of different model combinations and provided an in-depth review of the different performances and failure cases.