Control of a lower limb exoskeleton using Learning from Demonstration and an iterative Linear Quadratic Regulator...

Area of Science:

• Robotics
• Biomechanics
• Control Systems

Background:

• Lower limb exoskeletons require complex control to mimic human motion like walking.
• Existing learning from demonstration methods are limited by their use with linear quadratic regulators, unsuitable for non-linear robotic dynamics.

Purpose of the Study:

• To develop and validate a novel framework for real-time control of lower limb exoskeletons.
• To enable exoskeletons to accurately replicate human motion through advanced control signal optimization.

Main Methods:

• Utilized an Asynchronous Multi-Body Framework for simulating exoskeleton dynamics and enabling real-time control.
• Recorded eleven gait cycle demonstrations from volunteers using motion capture.
• Encoded demonstrations using Task Parameterized Gaussian mixture models.
• Employed an iterative linear quadratic regulator to optimize control signals for desired joint trajectories.
• Integrated a PD controller, optimized via Bayesian Information Criterion, for unmodeled dynamics.

Main Results:

• Successfully learned exoskeleton joint trajectories from human motion demonstrations.
• Optimized control signals by reducing learning bins, demonstrating efficient trajectory replication.
• The presented framework generated optimal control signals for accurate human motion following.

Conclusions:

• The developed framework effectively addresses the challenge of controlling non-linear dynamics in lower limb exoskeletons.
• This approach enhances the ability of exoskeletons to precisely mimic human gait, advancing assistive and rehabilitative technologies.