11/09/2026
By Thomas Dugan

The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a doctoral dissertation defense by Stephanie Epstein, titled “Development of a Unified Outer-Loop Control Strategy for Improved Actuation in Active Lower-Limb Exoskeletons.”

Details

  • Date: Friday, Nov. 20, 2026
  • Time: 10 a.m. – noon
  • Location: Lydon 203

Committee members

  • Murat Inalpolat, Advisor, Professor, Mechanical and Industrial Engineering, UMass Lowell
  • Paul Robbinette, Associate Professor, Electrical and Computer Engineering, UMass Lowell
  • Kshitij Jerath, Associate Professor, Mechanical and Industrial Engineering, UMass Lowell
  • Yi-Ning Wu, Associate Professor, Physical Therapy and Kinesiology, UMass Lowell
  • Pei-Chun Kao, Associate Professor, Physical Therapy and Kinesiology, UMass Lowell

Abstract

Active lower-limb exoskeletons must respond to irregular and abrupt changes in human gait, including variations caused by uneven terrain, unexpected obstacles, and rapid transitions between movement patterns. Although current systems can provide reliable assistance during a limited range of tasks, their ability to adapt in real time is constrained by processing demands and decision-making requirements.

This dissertation develops computationally efficient models and a unified control strategy for active lower-limb exoskeletons using limited sensing. A single-input, single-output long short-term memory model is developed to forecast gait-related transitions and provide the information needed for real-time actuation. The model is incorporated into a unified outer-loop control strategy designed to support both transition and steady-state tasks while communicating actuation commands to the device’s low-level controller.

The strategy is evaluated using data from five human subjects and across three test scenarios involving uncontrolled terrain, controlled walking, and uncontrolled speed. To reduce prediction latency, the controller is further enhanced through an online-learning model based on seasonal autoregressive time-series methods. Although the online model produces a modest increase in prediction error compared with the LSTM model, it offers substantially faster computational performance.

The results indicate that online learning can provide a practical alternative to higher-latency deep-learning models for detecting gait transitions and supporting real-time exoskeleton control. The dissertation establishes a foundation for lower-latency timing estimation, assistance across multiple movement scenarios, and future autonomous operation of active lower-limb exoskeletons outside highly controlled laboratory environments.