International audience This paper addresses the problem of synthesizing in real time the motion of realistic virtual characters with a physics-based model from the analysis of human motion data. The synthesis is achieved by computing the motion equations of a dynamical model controlled by a sensory motor feedback loop with a non-parametric learning approach. The analysis is directly applied on end-effector trajectories captured from human motion. We have developed a Dynamic Programming Piecewise Linear Approximation model (DPPLA) that generates the discretization of these 3D Cartesian trajectories. The DPPLA algorithm leads to the identification of discrete target-patterns that constitute an adaptive sampling of the initial end-point trajectory. These sequences of samples non uniformly distributed along the trajectory are used as input of our sensory motor system. The synthesis of motion is illustrated on a dynamical model of a hand-arm system, each arm being represented by seven degrees of freedom. We show that the algorithm works on multi-dimensional variables and reduces the information flow at the command level with a good compression rate, thus providing a technique for motion data in-dexing and retrieval. Furthermore, the adaptive sampling seems to be correlated with some invariant law of human motion.