Machine intelligence and automation have greatly increased the efficiency of the assembly industry. However, engineers have to design custom automation pipelines for different objects or object categories due to the lack of an algorithm with human-level intelligence that can assemble distinct objects. As a result, seasonal products are often still assembled manually. In this project, we aim to develop an algorithm that can assemble arbitrary objects based only on visual observations using deep learning. The algorithm will be practically useful in the assembly industry and significant to current robotic intelligence research. Our approach is to learn how to disassemble an already assembled object based on its visual geometry. If the algorithm successfully solves the disassembly problem, reverting the disassembly sequence will give a valid assembly sequence, thus solving the assembly task. To train our algorithm, we will leverage the Fusion 360 Gallery Dataset, a dataset recently released by the AutoDesk research team that comprises thousands of assembly models. We will train our algorithm in a simulated environment for efficiency and test our algorithm in the real world to validate its practicality.
Direct Supervisor: Shuran Song
Position Dates: Summer 2022
Hours per Week: 10
Number of positions: 1
Qualifications: Compter Graphics, Robotics
Eligibility: Sophomore, Junior, Senior