Biological Hip Torque Estimation using a Robotic Hip Exoskeleton.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Estimating joint kinetics is crucial for advanced exoskeleton control.
• Machine learning (ML) offers a promising approach for real-time kinetic estimation using limited sensor data.
• Current methods for estimating joint torque may not be optimal for dynamic, real-world applications.

Purpose of the Study:

• To evaluate the efficacy of ML models in estimating biological hip torque during various walking conditions.
• To compare ML-based torque estimation against a baseline mean torque profile method.
• To determine the suitability of these ML models for modulating exoskeleton assistance.

Main Methods:

• Calculated sagittal plane biological hip torque using inverse dynamics from human subject data.
• Developed and applied neural network (NN) and XGBoost ML models to estimate hip torque using only onboard exoskeleton sensor data.
• Compared ML model performance against a baseline method using the mean torque profile.

Main Results:

• NN and XGBoost models achieved significantly lower RMSE (0.116±0.015 and 0.108±0.011 Nm/kg) compared to the baseline (0.300±0.145 Nm/kg).
• ML models outperformed the baseline even after optimizing the baseline for specific ambulation modes.
• No significant difference in estimation accuracy was found between unified and separate ML models trained for different walking conditions.

Conclusions:

• ML algorithms can accurately estimate biological hip torque using only mechanical sensor data from a hip exoskeleton.
• These ML models show potential for real-time exoskeleton control and modulation of assistance.
• A single ML model can effectively estimate hip torque across different ambulation modes, simplifying implementation.