Learning to assemble furniture from their 2D drawings

Date

2023-12-07

Author

Uzel, Denge

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

319
views

235
downloads

Cite This

Prior work on learning furniture assembly assumes the availability of precise 3D information about the target furniture. This thesis elevates this assumption by learning to assemble furniture given a 2D drawing of its assembled form. To this end, the thesis introduces a novel network that can learn the similarity (conformity) between a 2D furniture drawing and a 3D point cloud representing the current state of the assembly. The proposed network is then used to formulate a reward signal for assembly learning using reinforcement learning. To ensure real-world applicability, a simulation environment generates a visually similar representation of the assembled furniture based on IKEA assembly instructions. The research encompasses three furniture classes: bookcase, chair, and table. A dedicated dataset is presented, including 2D furniture drawings resembling IKEA instructions and a 3D mesh model dataset encompassing various furniture assembly scenarios. The AssembleRL-2D model is trained using positive and negative input pairs from the 2D drawing and 3D mesh datasets, demonstrating proficiency across the three furniture classes. Notably, the model achieves accurate final furniture assembly, even in various assembly combinations where the parts of a chair model are assembled in different orders. AssembleRL-2D marks promising progress in furniture assembly learning, representing the inaugural application of a reinforcement learning model grounded in 2D final furniture assembly knowledge as a reward. The significance of this model in robotic assembly is highlighted by its capability to solve previously unencountered problems, showcasing its potential impact on addressing novel challenges in the field.

Subject Keywords

Assembly learning, 2D Manual-like image guided 3D furniture assembly, 2D drawing - 3D assembly similarity network, Reinforcement learning

URI

https://hdl.handle.net/11511/107815

Collections

Graduate School of Natural and Applied Sciences, Thesis