07/01/2026
By Brendan Hertel
Date: Wednesday, July 29
Time: 10-11 a.m.
Location: Perry Hall (PER) 415
Committee:
Reza Azadeh, Ph.D., Committee Chair, Associate Professor, UMass Lowell
Holly Yanco, Ph.D., Distinguished University Professor, UMass Amherst
Maria E. Cabrera, Ph.D., Assistant Professor, UMass Lowell
Harish Ravichandar, Ph.D., Assistant Professor, Georgia Institute of Technology
Abstract
Humans are able to adapt to the world around them effortlessly. They encounter rapidly changing environments, understand and accurately perform complex tasks, and need to synthesize new skills from known skills. Robots, on the other hand, struggle to achieve the same capability. Robots are often used in highly controlled environments, rely on previous environment knowledge, and are slow to adapt to new situations. To address these issues, we propose a novel end-to-end framework for robot learning, known as SERGE (Segmentation, Encoding, Recall, Generalization, and Execution). Our framework includes several novel components, and presents a practical integration of these components. Our framework first segments demonstrated tasks, allowing for learning from simple or complex demonstrations. Next, skills are clustered and encoded in a library of motion primitives for efficient reuse. When skills are needed for execution, we can recall these encoded skill models and generalize them to the current task and environment. We use a novel Learning from Demonstration approach to generalize known skills. Finally, we incorporate feedback into our model, allowing our framework to learn from itself and provide better executions in lifelong learning scenarios. The main contributions of this work include novel algorithms for segmentation, clustering, and encoding demonstrations, a novel method for Learning from Demonstration with feedback, and the conceptualization, implementation, and validation of the SERGE framework.