Deep-Learning Control of Lower-Limb Exoskeletons via Simplified Therapist Input

Area of Science:

• Rehabilitation Engineering
• Robotics
• Biomechanics

Background:

• Partial-assistance exoskeletons aid gait rehabilitation by encouraging active user participation.
• Current hierarchical control methods require extensive calibration and user-specific tuning, limiting adaptability for complex movements like stair navigation.

Purpose of the Study:

• To propose and evaluate a novel three-step, data-driven approach for exoskeleton control, overcoming limitations of traditional hierarchical methods.
• To enable intuitive adjustment of gait parameters by therapists and predict optimal exoskeleton assistance.

Main Methods:

• Probabilistically inferring locomotion states (e.g., step length, velocity, gait phase) from sensor data.
• Utilizing a user interface for therapists to modify inferred gait features.
• Predicting desired joint posture and spring-damper system stiffness based on adjusted features and prediction uncertainty.

Main Results:

• The data-driven approach demonstrated adaptability to varying walking and stair-climbing conditions in healthy participants.
• Negative interaction power at hip and knee joints indicated effective exoskeleton assistance across different gait characteristics.
• Kinematic variations correlated with adjusted gait features, validating the system's responsiveness.

Conclusions:

• The proposed data-driven method offers a more adaptable and less calibration-intensive alternative to hierarchical control for partial-assistance exoskeletons.
• This approach facilitates personalized gait rehabilitation by allowing therapist-guided adjustments to exoskeleton assistance.
• The findings support the potential of this method for enhancing gait (re)learning and rehabilitation outcomes.