The study presented in this paper aims at achieving optimal paths for a human-performed assembly task. Optimality, from an operator's point of view, means that minimal fatigue occurs. Trajectories for assembly parts, generated by potential fields, are optimized in terms of the work the operator has to do. Optimization is carried out by genetic algorithm. The trajectories have to satisfy the constraints imposed by physical human limitations. We therefore introduce a virtual human operator in the working environment, and determine by means of inverse kinematics if the trajectories are feasible. Our method is tested on the assembly of a mechanical part in 3D space.