2026-05-20T18:57:11Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2015002025-05-12T19:35:22Zoai_pmh:alloai_pmh:repo_items

201500 https://hdl.handle.net/2286/R.2.N.201500 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 48 pages Masters Thesis Academic theses en Liu, Yu-Lun Gopalan, Nakul Ben Amor, Hani Sanneman, Lindsay Arizona State University Partial requirement for: M.S., Arizona State University, 2025 Field of study: Computer Science This thesis investigates robotic manipulation in kitchen environments, focusing on imitation learning (IL) using Action Chunking with Transformers (ACT) and continual learning via ACT integrated with Low-Rank Adaptation (LoRA). The ACT framework is evaluated on 15 distinct kitchen tasks, focusing on single-arm operations such as Deformable Object Handling, Precision Grasping and Alignment, Force and Motion Control, and Dynamic Manipulation. Collected 20 human demonstrations for each of the 15 kitchen manipulation tasks. For generalization evaluation, 50 demonstrations were collected in three positional variations. In total, 450 demonstrations were collected in this thesis. Furthermore, ACT-LoRA was evaluated to examine its continual learning performance across multiple kitchen tasks. Experimental results reflect the capabilities of ACT and identify limitations, including challenges with non-Markovian behavior, millimeter-level precision, task-specific generalization, and physical interactions. Potential future directions involve addressing these limitations and exploring efficient techniques for managing non-Markovian behavior in imitation learning. Computer Science Robotics Multi-Task Continual Learning in Robotics for Cooking