Gaussian Process Trajectory Learning and Synthesis of Individualized Gait Motions.

Area of Science:

• Robotics
• Machine Learning
• Biomechanics

Background:

• Robotic gait rehabilitation requires accurate and adaptable reference motion generation.
• Existing methods struggle with individualized gait patterns and arbitrary speed adjustments.

Purpose of the Study:

• To develop a Gaussian process-based method for learning and generating individualized gait trajectories.
• To enable the generation of reference motions for robotic gait rehabilitation systems at user-defined walking speeds.

Main Methods:

• Utilized nonlinear dimension reduction with Gaussian process dynamical models (GPDMs) for gait dynamics.
• Modeled internal dynamics as a second-order Markov process in a lower-dimensional latent space.
• Employed Gaussian process regression (GPR) to predict latent space dynamics based on desired speed and anthropometrics.

Main Results:

• Generated gait patterns with 30% less mean square prediction error compared to state-of-the-art methods.
• Successfully generated individualized gait motions at arbitrary user-specified walking speeds.
• Experimental validation with over 100 subjects confirmed method efficacy.

Conclusions:

• The proposed Gaussian process-based method offers a robust approach for generating personalized gait motions.
• This technique enhances the adaptability and accuracy of robotic gait rehabilitation systems.
• The method's ability to handle diverse speeds and subject parameters represents a significant advancement.