Gait Trajectory and Gait Phase Prediction Based on an LSTM Network.

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Lower body segment trajectory and gait phase prediction are essential for controlling robotic assistance devices like exoskeletons.
• Accurate prediction enables powered exoskeletons to provide timely and appropriate assistance during gait.

Purpose of the Study:

• To develop and validate a Long Short-Term Memory (LSTM) network for predicting lower body segment trajectories and gait phases.
• To enhance the control capabilities of assistance-as-needed robotic devices by compensating for system delays.

Main Methods:

• Utilized inertial measurement units (IMUs) on thigh, shank, and foot segments to collect data.
• Developed a modified LSTM network, a type of recurrent neural network, for sequence prediction.
• Implemented a weighted discount loss function to prioritize short-term prediction accuracy (3-5 frames) while maintaining overall performance (up to 10 frames).

Main Results:

• The LSTM model achieved high correlation (r > 0.98) between predicted and measured lower limb segment trajectories.
• Accurate prediction of five distinct gait phases was demonstrated, with 95% accuracy for the swing phase in inter-subject implementation.
• The model showed strong generalization capabilities across different participants.

Conclusions:

• The proposed LSTM approach effectively predicts future gait trajectories and phases, crucial for real-time exoskeleton control.
• This predictive capability can significantly improve exoskeleton controller design, leading to smoother gait phase transitions and enhanced wearer assistance.
• The findings pave the way for more responsive and intuitive exoskeleton-human interaction.