Year: 2025
Type: dissertation
Abstract: Robotic manipulation at the microscale has gained significant relevance in fields such as biomedical engineering, micro-assembly, and cellular manipulation. Despite advancements in robotic micromanipulators, autonomous control at the microscale remains a largely open area of research, particularly in the context of dexterous manipulation of soft objects. Such manipulation requires high precision and adaptability, posing challenges for traditional control methods due to the difficulty in modeling the behavior of soft objects. This thesis presents the development of a dual robotic microgripper capable of autonomous, dexterous micromanipulation of soft microfibers, demonstrated through shaping tasks. To enable this, learning-from-demonstration control strategies, specifically behavioral cloning, are employed. By learning from expert demonstrations, the system infers the dynamics of soft microfibers without requiring an explicit model. Two control algorithms are implemented to achieve the shaping task: a K-Nearest Neighbor (KNN) regressor and a trained neural network (NN) policy. Their performance is evaluated through two experiments. The first involves repeated manipulation between fixed initial and target states. The second requires manipulation from multiple initial states, including those not present in the training data, to a fixed target state. In the first experiment, the KNN and NN approaches achieved positioning accuracies and precisions of 27 ± 1 µm in 15 seconds and 29.2 ± 3.2 µm in 8 seconds, respectively. In the second experiment, they reached accuracies and precisions of 570 ± 291 µm in 7.5 seconds and 231 ± 34 µm in 4.2 seconds, respectively. The KNN method exhibited high sensitivity to demonstration quality, while the NN demonstrated the ability to generalize across unseen states, indicating greater adaptability.
Source: Aaltodoc (Aalto University)
Author: Jaakkola, Veeti
Institution: Aalto University
Language: English
Topic: Micro and Nano Robotics
Subfield: Condensed Matter Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7076961400
