The demand for realistic autonomous virtual humans is increasing, with potential application to prototype design and analysis for a reduction in design cycle time and cost. In addition, virtual humans that function independently, without input from a user or a database of animations, provide a convenient tool for biomechanical studies. However, development of such avatars is limited. In this paper, we capitalize on the advantages of optimization-based posture prediction for virtual humans. We extend this approach by incorporating multi-objective optimization (MOO) in two capacities. First, the objective sum and lexicographic approaches for MOO are used to develop new human performance measures that govern how an avatar moves. Each measure is based on a different concept with different potential applications. Secondly, the objective sum, the min-max, and the global criterion methods are used as different means to combine these performance measures. It is found that although using MOO to combine the performance measures generally provides reasonable results especially with a target point located behind the avatar, there is no significant difference between the results obtained with different MOO methods.