- Marie Dominique Schmidt1 , Ioannis Iossifidis1
- Institute of Computer Science, University of Applied Science Ruhr West, Duisburger Str. 100, 45479 Mülheim an der Ruhr, Germany
The upper limbs enable us to perform a variety of tasks in everyday life that require strength and a wide range of motion as well as precision. For coordinated motion, the action must be well planned and timed. Therefore, information about the target and the current body state from the sensory systems is as important as the integration of previous experiences. These experiences might be represented as pre-learned inverse dynamics that generate associated muscle activity. We propose a generative model that predicts the upper limb muscle activity driven by various motion parameters. The generative model is based on a recurrent neural network predicting muscle activity or complex upper limb motions. Our approach achieves remarkable agreement in predicting different subjects and abstracts well for new motions and muscle activities that have not been trained before. The high inter-subject variation of the recorded muscle activity is successfully handled using a transfer learning approach, resulting in a good fit for a new subject.
To gain a deeper understanding of the link between motion trajectory and muscle activity, we reverse our problem and test the opposite prediction from muscle activity to motion parameters. The ability of this approach to predict muscle activity and motion trajectory efficiently has implications for the fundamental understanding of movement control and use for rehabilitation of neuromuscular diseases with myoelectric prostheses and functional electrical stimulation.