A diagnostic scheme for actuator and sensor faults which can occur on a robot manipulator using a model-based fault diagnosis (FD) technique is addressed. With the proposed FD scheme, it is possible to detect a fault, which can occur on a specific component of the system. To detect actuator faults, higher order sliding-mode unknown input observers are proposed to provide the necessary analytical redundancy. The detection of sensor faults, instead, is made by relying on a generalized observer scheme. The observer input laws are designed according to two well-known second-order sliding-mode approaches: the so-called supertwisting and the suboptimal one. Both typologies of input laws allow to perform a satisfactory FD. The peculiarities of each input law of the observers are discussed. To make possible fault isolation, it is required that a single fault acts only on one component of the system at a time. If one knows that faults occurred only on actuators, then it is possible to isolate multiple simultaneous faults on actuators. The proposed approach is verified in simulation and experimentally on a COMAU SMART3-S2 robot manipulator. © 2012 IEEE.