Robot-assisted therapy contingent on the intention to move ensures the active engagement of patients during training. However, detecting the intention to move in severely impaired patients with no visible movement is a challenge where physiological signals such as Electromyogram signals (EMG) can be used. An effective EMG-driven robot-assisted therapy in severely impaired patients should provide naturalistic human-machine interaction, which requires an optimal EMG detector with high detection accuracy and low latency. Non-availability of ground truth about the presence/absence of EMG in severely impaired patients with no movement is a challenge, which hinders the computation of detection latency and accuracy. Therefore, this paper identifies an optimal EMG detector without the ground truth about the presence of residual EMG signals. An unsupervised approach using total variation distance was used for this purpose to distinguish between the rest state when the muscle is fully relaxed and the move state where there could be muscle activity. The analysis was done on residual EMG data from one severely impaired stroke patient. The results reveal that the modified Hodges and approximate generalized likelihood ratio (AGLR) detectors maximally separate the rest and move states. The AGLR detector showed poor performance in both detection latency and accuracy, whereas the modified Hodges detector demonstrated better performance, making it a potentially better choice for personalized EMG-driven robot-assisted therapy.