An algorithm to reduce human-robot interface compliance errors in posture estimation in wearable robots

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Accurate human posture estimation is crucial for wearable robot controllers.
• The compliance of human-robot interfaces can introduce significant errors in posture estimation.
• Existing methods struggle with the dynamic and compliant nature of human-robot interaction.

Purpose of the Study:

• To develop and validate a machine learning algorithm for correcting posture estimation errors in wearable robots.
• To improve the accuracy of thigh segment angle estimation for wearable robot control.
• To assess the algorithm's effectiveness across various walking speeds and assistance levels.

Main Methods:

• Collected motion capture and wearable robot data from 8 participants walking on a treadmill.
• Used optical motion capture to measure relative displacement between user and robot (Myosuit).
• Trained a gradient boosting model (XGBoost) using combined user and robot data to correct for mechanical compliance errors.

Main Results:

• Reduced root mean square error (RMSE) of thigh segment angle estimation from 6.3° to 2.5°.
• Decreased average maximum error in thigh angle estimation from 13.1° to 5.9°.
• Observed significant improvements across all tested assistance force levels and walking speeds.

Conclusions:

• Machine learning offers a promising solution for accurate user posture estimation in wearable robots.
• The developed algorithm effectively corrects for mechanical compliance errors.
• Enhanced posture estimation can improve the performance and safety of assistive wearable robots.