A patient-driven control strategy for Functional Electrical Stimulation (FES), which amplifies volitionally-initiated shoulder abductions, is proposed to improve stroke patients' rehabilitation. Based on the measured abduction angle, a FES-induced muscle recruitment is generated that yields a pre-specified percentage of this angle - yielding arm weight relief. To guarantee the correct recruitment also under fatigue and uncertain muscle activation we employ feedback control of the recruitment level determined by filtering the FES-evoked electromyogram. Filter parameters are user-optimized to obtain a linear relation between filter output and angle with a good signal-to-noise ratio. The auto-tuned recruitment controller (RC) was tested on five healthy subjects and compared to direct stimulation (DS) while muscle fatigue progressively occurred. Results showed a more linear relation between recruitment level and angle than between non-controlled stimulation intensity and angle (R(2)=0.93 vs. R(2)=0.79, angular range of 54°). After 6 min of stimulation, abduction decreased by 42% ± 14 for DS and by 0% ± 12 for RC, showing an effective compensation of fatigue. RC yielded significant smaller errors than DS in generating desired angles (0.23% ± 5.9 vs. 14.6% ± 9.7). When FES-induced arm weight support was provided, a mean reduction of the volitional effort (determined by Electromyography) of 78% was achieved compared to angular tracking without FES. First experiments with one acute stroke patient are also reported.