Mechanical condition monitoring of robots can be used to detect unexpected failure of robots. Data taken from a robot operation is processed and compared against a health baseline. Features extracted during the monitoring stage of robot operation are aligned with features extracted during the training stage in which the health baseline is established by projecting both onto a common subspace. A classifier which can include a distance assessment such as an L2-norm is used within the common subspace to assess the condition of the robot. Excursions of the distance assessment from a criteria indicate a failure or potential failure.