The generation of believable human-like movements is a key problem of humanoid robotics. A central biologically-inspired approach for the solution of the underlying high-dimensional control problems relies on learned low-dimensional dynamic movement primitives. METHOD: Three different control algorithms, based on different definitions of dynamic movement primitives, were trained with human motion-capture data from a double-step reaching task. The algorithms were applied to control an accurate physical model of the humanoid robot COMAN (developed at the Istituto Italiano di Tecnologia). The generated movements were then used to animate a corresponding robot avatar model. In addition, the original human movements were retargeted to the same avatar model. Participants rated the ‘human-likeness’, the ‘naturalness’, and reported observed ‘artifacts’. RESULTS: Interestingly, participants rated not the human but the most smooth trajectories as most ‘natural’ and most ‘human-like’. This points to a non-veridical perception of human-likeness, which might be based rather on low-level properties of the observed motion than on the reproduction of details of complex human trajectories. [ This research was supported by: AMARSi- EC FP7-ICT-248311; Koroibot FP7-ICT-2013-10/ 611909; DFG GI 305/4-1, DFG GZ: KA 1258/15-1; BMBF, FKZ: 01GQ1002A, FP7-PEOPLE-2011-ITN(Marie Curie): ABC PITN-GA-011-290011, HBP FP7-ICT-2013-FET-F/ 604102. ]