Generating Real-time Muscle Activations for Skeletal Hand Motion: An Evolutionary Approach

This paper presents a system that gener- ates muscle activations from captured motion by bor- rowing principles from biomechanics and neural con- trol. A physics engine compliant muscle system is de- veloped using the human hand as an example. A ma- chine learning approach using evolutionary neural net- works is adopted for creating the muscle control system and dynamical simulation is performed using a real- time physics engine that is used in present day games. The system was trained for a single nger and then tested on the ve ngers of the hand. The simulation results show that the system produced a close mimic of the motion-captured hand animation. The system also produced a believable animation from motion data not present in the training set.