Artificial neural networks for motion emulation in virtual environments

Abstract

Simulation of natural human movement has proven to be a challenging problem, difficult to be solved by more or less traditional bioinspired strategies. In opposition to several existing solutions, mainly based upon deterministic algorithms, a data-driven approach is presented herewith, which is able to grasp not only the natural essence of human movements, but also their intrinsic variability, the latter being a necessary feature for many ergonomic applications. For these purposes a recurrent Artificial Neural Network with some novel features (recurrent RPROP, state neurons, weighted cost function) has been adopted and combined with an original pre-processing step on experimental data, resulting in a new hybrid approach for data aggregation. Encouraging results on human hand reaching movements are also presented.