Intelligent control of exoskeletons through a novel learning-from-demonstration method

Ugur, E.; Samur, E.; Ugurlu, B.; Erol Barkana, D.E.; Kucukyilmaz, A.; Bebek, O.

Abstract

We present a novel concept that enables the intelligent and adaptive control of exoskeletons through exploiting our state-of-the-art learning from demonstration (LfD) method, namely Conditional Neural Movement Primitives (CNMPs) [1], on our integrated system of a soft suit and a robotic exoskeleton [2]. Learning of complex locomotion trajectories is aimed to be achieved first by learning from demonstrated trajectories of healthy walking human subjects, next by refining the skills using Reinforcement Learning (RL) methods taking into account the balance and energy consumption of the exoskeleton system, and finally by realizing feedback-based control that enables robust execution in the face of external unexpected perturbations. To the best of our knowledge, LfD and RL based exoskeleton trajectory control is new in the field where existing ones are limited to trajectory generation without taking into account the exoskeleton-human interaction dynamics. Our LfD method can extract the prior knowledge directly from the training data by sampling observations, and uses it to predict a conditional distribution over any other target points. CNMPs specifically learn complex temporal multi-modal sensorimotor relations in connection with external goals, produce movement trajectories, and execute them through a high-level feedback control loop.

To react to unexpected events during action execution, CNMP can be conditioned with sensor readings in each time-step. To transfer learned knowledge from human demonstrations to control of our exoskeleton that has different physical properties, we extend CNMPs with state-of-the-art RL methods emphasizing generalization.

Citation

Ugur, E., Samur, E., Ugurlu, B., Erol Barkana, D., Kucukyilmaz, A., & Bebek, O. (2020, September). Intelligent control of exoskeletons through a novel learning-from-demonstration method. Poster presented at Cybathlon Symposium 2020, Zurich, Switzerland

• Presentation Conference Type: Poster
• Conference Name: Cybathlon Symposium 2020
• Start Date: Sep 17, 2020
• End Date: Sep 18, 2020
• Deposit Date: Mar 20, 2022
• Publicly Available Date: Mar 21, 2022
• Keywords: Exoskeletons, control, machine learning, learning from demonstration, reinforcement learning
• Public URL: https://nottingham-repository.worktribe.com/output/7641578
• Publisher URL: https://cybathlon-symposium.ethz.ch/