This paper presents methods for generating optimal reach-ing and balancing movements in humanoid robots. Past approaches to dynamics-based optimal robot movement gen-eration, e.g., [1], [7], [4], have been computationally pro-hibitively expensive for online generation. To address thesecomputational difficulties, attempts to emulate the humancapability of storing and reusing movement modules havealso been made in different contexts throughout the roboticsand human motor control literature. The idea of representingand generating human arm motions as a linear superposi-tion of principal components was first investigated in [13].Similar approaches using human motion data have also beeninvestigated for deriving action and behavior primitives forhumanoid robots [3], [9], although these approaches are ex-clusively kinematics-based. It should also be noted that theoptimal movements generated offline using any of the previousdynamics-based optimization schemes can in principle be usedas primitives for schemes such as the one presented in thispaper. To address the optimal reaching problem, various arm reaching movements are parametrized as a linear combinationof basis elements, which are extracted via a principal com-ponent analysis of human motion capture data. Human-likereaching movements are then generated by determining theminimum torque motion from this parametrized class of armmovements.