Accurate Real-Time Joint Torque Estimation for Dynamic Prediction of Human Locomotion.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Traditional inverse dynamics for joint torque estimation requires both kinematic and kinetic data.
• Real-time analysis of human movement is crucial for applications like prosthetics and rehabilitation.

Purpose of the Study:

• To develop and evaluate novel real-time methods for estimating human joint torques using only kinematic data.
• To compare the performance of a hybrid model and an end-to-end neural network against traditional methods and each other.

Main Methods:

• A novel real-time hybrid method integrating a neural network and a dynamic model was developed.
• An end-to-end neural network for direct joint torque estimation was also created, both relying solely on kinematic data.
• Models were trained on diverse walking conditions and validated using simulations (OpenSim) and experimental data, including with a lower limb exoskeleton.

Main Results:

• The hybrid model achieved high accuracy in simulations (RMSE < 5 N.m, R > 0.95).
• The end-to-end model outperformed the hybrid model when compared to gold-standard methods requiring kinetic and kinematic data.
• However, the hybrid model demonstrated superior performance (R=0.84) over the end-to-end network (R=0.59) when tested on a participant wearing an exoskeleton, indicating better generalization.

Conclusions:

• Kinematic-only methods offer a viable alternative for real-time joint torque estimation.
• The hybrid model shows greater robustness and applicability in scenarios deviating from training data, such as with exoskeleton use.
• Further research can refine these models for broader clinical and robotic applications.