04/24/2026
By Danielle Fretwell

The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a Doctoral Dissertation Proposal defense by Stephanie Epstein titled: "Development of Realtime Decision-making Strategies for Improved Actuation Accuracy of Active Lower Limb Exoskeletons"

Candidate Name: Stephanie Epstein
Degree: Doctoral
Defense Date: Friday, May 8, 2026
Time: 9 - 11 a.m.
Location: Southwick 240

Committee:

  • Advisor: Murat Inalpolat, Professor, Mechanical and Industrial Engineering, University of Massachusetts Lowell
  • Paul Robinette, Associate Professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • Kshitij Jerath, Associate Professor, Mechanical and Industrial Engineering, University of Massachusetts Lowell
  • Yi-Ning Wu, Associate Professor, Physical Therapy and Kinesiology, University of Massachusetts Lowell
  • Pei-Chun Kao, Associate Professor, Physical Therapy and Kinesiology, University of Massachusetts Lowell

Brief Abstract:
While actively powered exoskeletons effectively provide actuation to assist users for their repetitive and more conventional steady-state tasks, they generally fall short of assisting with transitional tasks comprising sudden changes in speed, task or terrain type. This is predominantly caused by inaccuracies both in actuation magnitude and actuation timing, leading to declined assistance provided by the exoskeleton. When proper assistance is provided by the exoskeleton, the user’s overall metabolic cost will usually lessen, allowing one to continue to perform an activity for a longer period of time and with higher accuracy. A significant amount of recent research has been on advancing exoskeleton control strategies for specific test scenarios such as steady-state walking, running on a treadmill or on an inclined surface, or climbing stairs at a relatively constant speed. However, these strategies are expected to help controllers adapt fast and make swift decisions for their real-time implementation and assistance with fast-paced transitional tasks. This requires inferring user intent and using this information to make informed decisions in real-time regardless of the activity or its pace. A series of machine learning based adaptive decision-making strategies is proposed to help with the real-time controls of lower-limb exoskeletons for their improved actuation accuracy for both steady-state and transitional movements. The proposed strategies are accounting for the abrupt changes in human locomotion as well as the upcoming task pattern and thus are expected to help the controller to adapt and accurately augment human locomotion while increasing task accuracy and performance. The overall research goal is to create strategies based upon human intention while using limited sensors for feedback keeping the overall system computationally and structurally light. Current research barriers exist due to mobile computers’ latency and insufficient memory or space to process the information quickly where the controller ends up falling behind the user’s actions. Additionally, implementing and testing these controllers and making informed decisions in real-time while using machine learning techniques has been challenged in a setting which requires mobility. The currently available control algorithms originally designed and implemented on the commercially available active lower limb exoskeleton used in the current study only assists users during steady walking. As soon as one deviates from a constant paced walking pattern, this controller stops providing additional assistance at its current state. This doctoral dissertation research proposes the design, testing and validation of adaptive decision-making strategies for active exoskeleton controllers enabling the exoskeletons being not only helpful with steady-state tasks conducted on a treadmill but also with transitional tasks experienced in various environments including different terrain types. The proposed work entails building efficient machine learning algorithms based upon collected human subject data from different terrains, changes of speeds of the tasks and obstacles experienced while on these terrains. Additionally, the proposed algorithms are tailored to help the controller make decisions based upon limited sensor feedback while maintaining quick controller iteration timing. These uniquely efficient algorithms are implemented into an active lower limb exoskeleton and will be used for the validation of the improved real-time controls strategy. Demonstration of the proposed algorithms’ impact on the actuation accuracy will be achieved through human subject testing.