Robot learning and planning for autonomous disassembly
Loading...
Authors
Draper, James, Austin
Issue Date
2025-12
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Alternative Title
Abstract
Robotic disassembly is a common approach to efficiently implement recycling at scale. Methods exist for perception and manipulation to autonomously perform the disassembly of var- ious products. However, the majority of these methods focus on planning disassembly sequences given the product’s construction and component relationships are known. We propose methods for learning graphical representations of products and planning disassembly sequences in real time. We extend the existing problems to scenarios where product assembly information is not explicitly known. Given observed features of a product, we demonstrate the use of neural networks for predicting the component precedence relationships as AND/OR graphs. These graphs inform action-availability for online planning methods for real-time disassembly. We propose Monte Carlo Tree Search as a robust method for real-time disassembly planning when visual obstructions cause a partial view of product features. The proposed framework is evaluated with real-time simulations, demonstrating the learned graphical models and online planning approach can perform effective disassembly without full prior knowledge of the product structure. Simulations show a success rate of 99.9%, with an average disassembly plan length of 1.49 times the optimal number of component removals. These results suggest a promising direction for scalable, autonomous recycling systems capable of handling diverse and unstructured products.
Description
December2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY