Most recent findings in robot-assisted therapy suggest that the therapy is more successful if the patient actively participates to the movement (“assistance-as-needed”). In the present contribution, we propose a novel approach for designing highly flexible protocols based on this concept. This approach uses adaptive oscillators: a mathematical primitive having the capacity to learn the high-level features of a quasi-sinusoidal signal (amplitude, frequency, offset). Using a simple inverse model, we demonstrate that this method permits to synchronize with the torque produced by the user, such that the effort associated with the movement production is shared between the user and the assistance device, without specifying any arbitrary reference trajectory. Simulation results also establish the method relevance for helping patients with movement disorders. Since our method is specifically designed for rhythmic movements, the final target is the assistance/rehabilitation of locomotory tasks. As an initial proof of concept, this paper focuses on a simpler movement, i.e. rhythmic oscillations of the forearm about the elbow.