Insights into motor control: predict muscle activity from upper limb kinematics with LSTM networks.

This study explores the relationship between upper limb kinematics and corresponding muscle activity, aiming to understand how predictive models can approximate motor control. We employ a Long Short-Term Memory (LSTM) network trained on kinematic end effector data to estimate muscle activity for eight muscles. The model exhibits strong predictive accuracy for new repetitions of known movements and generalizes to unseen movements, suggesting it captures underlying biomechanical principles rather than merely memorizing patterns. This generalization is particularly valuable for applications in rehabilitation and human-machine interaction, as it reduces the need for exhaustive datasets. To further investigate movement representation and learning, we analyze the impact of motion segmentation, hypothesizing that breaking movements into simpler components may improve model performance. Additionally, we explore the role of the swivel angle in reducing redundancy in arm kinematics. Another key focus is the effect of training data complexity on generalization. Specifically, we assess whether training on a diverse set of movements leads to better performance than specializing in either simple, single-joint movements or complex, multi-joint movements. The study is based on an experimental setup involving 23 distinct upper limb movements performed by five subjects. Our findings provide insights into the interplay between kinematics and muscle activity, contributing to motor control research and advancing neural network-based movement prediction.