The nonlinearity and the time-varying parameters of the muscle models represent serious obstacles for the closed loop control of FES (Functional Electrical Stimulation). Different standard closed loop approaches, like PID, are known to be unreliable. In this paper, an adaptive control system with nonlinear compensation of the muscle recruitment curve is presented. The approximate correction of the recruitment curve is shown to be a necessary condition to insure control reliability. The same holds for the adaptation switching mechanism and for the linear forgetting algorithm. Experimental results for the control of the free swinging knee joint of able bodied persons in sitting and lying position are presented, showing the effectiveness of the proposed approach.