Year: 2025
Type: dissertation
Abstract: This thesis presents a transfer learning framework designed to enhance the data efficiency and generalizability of learning the kinematics of robots. The framework leverages pre-training on simulated datasets to significantly reduce the amount of real-world data required to achieve high accuracy in both forward kinematics and inverse kinematics modeling. The effectiveness of this approach was demonstrated through comprehensive evaluations on concentric tube continuum robots. For forward kinematics, the framework achieved sub-millimeter accuracy using only 10% of the real-world data needed by a baseline model trained from scratch. For inverse kinematics, it reduced real-world data requirements by 33% while maintaining sub-10 degree accuracy in rotational joint space errors. These results highlight the framework’s capacity to effectively utilize simulated data, ensuring robust pre-trained models that require minimal fine-tuning with real-world data. The framework also demonstrated strong generalizability, particularly in its application to different robot designs. For forward kinematics, the framework achieved comparable accuracy with a 75% reduction in real-world data, while for inverse kinematics, a 33% reduction was observed. These findings underscore the robustness of the transfer learning approach, which adapts well to variations in robot configurations with minimal additional data. Beyond concentric tube continuum robots, the principles underlying this transfer learning framework have broader applicability across various robotic systems. By reducing the reliance on extensive real-world data, this approach accelerates the development of accurate and reliable robotic models, making advanced robotic technologies more accessible and flexible for real-world applications.
Source: TSpace (University of Toronto)
Author: Chen, Ryan Zeyuan
Institution: University of Toronto
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7132934114
