Pathological tremor constitutes the most common movement disorder, and is increasing its prevalence with ageing. Treatment forms range from drugs to surgery in those patients refractory to drugs, however, tremor is not effectively managed in about 25% of patients. According to this, new management techniques such as wearable robots that take advantage of selective biomechanical loading seem an interesting alternative. Our objective is to design robotic exoskeletons which suppress tremor, letting the user perform a voluntary movement, by means of intelligent control approaches that include accu-rate tremor models. In this context, we propose a two-stage algorithm for real-time estimation of time varying tremor amplitude and frequency. It is based on the assumption that tremor alters voluntary motion in an additive manner, and happens in a higher frequency band. The two-stage algorithm first generates an estimation of voluntary movement based on its inherent slower dynamics, and then removes it from the total motion, directly providing an estimate of tremor. This tremor estimation is then fed into an adaptive filter, which provides instantaneous tremor characteristics. Accurate and robust tremor amplitude and frequency estimates are obtained.