A comparative study of ANN-based forward dynamics and inverse dynamics in human gait analysis

Area of Science:

• Biomechanics
• Robotics
• Computational Neuroscience

Background:

• Traditional inverse dynamics analysis of human motion relies on measured ground reaction forces and kinematic data.
• Forward dynamics simulations offer a more direct approach but require accurate models of muscle activation and control.
• Artificial Neural Networks (ANNs) show promise in developing sophisticated controllers for forward dynamics simulations.

Purpose of the Study:

• To compare the accuracy of human walking motion analysis between traditional inverse dynamics and an ANN-based forward dynamics method.
• To evaluate the ANN controller's ability to reproduce accurate gait kinematics and joint torques.
• To assess differences in mechanical energy consumption estimation between the two dynamic modeling approaches.

Main Methods:

• Collected motion capture and ground reaction force data from nine healthy male subjects during walking.
• Performed inverse kinematics and dynamics analysis using OpenSim.
• Trained an ANN-based gait controller using deep reinforcement learning in forward dynamics simulations, optimizing for kinematic tracking and reduced joint torques/power fluctuations.

Main Results:

• The ANN controller reproduced joint kinematics with a root-mean-square (RMS) difference of less than 2.7° compared to inverse kinematics.
• Joint torque profiles from the ANN method showed RMS differences of 0.20-0.23 Nm/kg, comparable to optimization-based methods.
• Forward dynamics with the ANN controller estimated higher total mechanical power consumption (underestimating by only 4.1% with residuals) compared to inverse dynamics (underestimating by 0.74 W/kg due to residual forces).

Conclusions:

• ANN-based forward dynamics modeling can accurately replicate human gait kinematics and joint dynamics.
• This approach offers a more accurate estimation of mechanical energy consumption by minimizing reliance on residual forces.
• The adaptable ANN controller has potential applications in analyzing gait variations for rehabilitation and developing assistive robotic devices.