Stroke patients have a decreased ability in performing activity of daily living (ADL) tasks such as in "drinking a glass of water", "lifting a bag", "turning a key" and so on. Sensorimotor force and torque measurements from patients performing these standardized ADL tasks are hypothesized to give quantitative information about the recovery process. Parts of the force/torque measurements contain useful information, when related to the initiation of the movement during ADL tasks. Here we address the challenging problem of automatically extracting the movement initiation from these force/torque measurements. We will adopt a machine learning approach which relies on the statistically rigorous maximal information redundancy (MIR) criterion. This assumes that movement initiation parts of the signals are characterized by an increased redundancy in the signal. A thorough evaluation of the criterion shows that the accuracy of the criterion in movement onset detection is close to that of clinical experts.