The need for making virtual humans more realistic keeps growing in domains like virtual reality, computer animation or interactive ergonomics. However, while many domains aim to enhance user's experience, applications in ergonomics mainly focus on the synthesis of realistic movements [1]. The number of works dealing with inverse kinematics to synthesize realistic movements highlights its difficulty. Among them, two categories can be sorted out of those works. On one hand, the whole inverse problem is learned from motion capture data and then reproduced [4]. On the other hand, a common method is used to solve the inverse problem and constraints are used to enhance realism. These constraints come from movement studies [2] or are extracted from captured data [5].To enhance the realism of the synthesized movements, we propose a new constraint model based on the representation of joint synergies. Our constraint model integrates into the sensorimotor model proposed by Gibet et al. [3]. This sensorimotor model, based on the Jacobian transpose method, includes a nonlinear gain function ("sigmoid" shape) to produce more natural pointing movements. We propose to replace this gain function by a new gain model learned from captured motions. After a description of our model, we present the learning process and the preliminary results that we obtained.We propose to model joint synergies as a vector of gains − → w , each w i being applied successively to the n degrees of freedom. This gain can be a function of time (t), current posture ( − → q t ), current angular speed (∆ − → q t ) and previous computed gain. It is computed at each step of the sensorimotor loop. The gain function may also use some constant parameters − → p to differentiate individuals.The size m of − → p is independant of the number n of degrees of freedom. These parameters may constitute the signature of an individual movements. We thus obtain the following model of joint synergies :The choice of the function f is discussed in the results but has no influence on the learning process of the parameters presented below. The goal is to find a value of − → p that makes possible the motion controller to reproduce recorded motions. We use a meta heuristic to adjust the parameters