Neurological patients with impaired upper limbs often receive arm therapy to restore or relearn lost motor functions. During the last years robotic devices were developed to assist the patient during the training. In daily life the diversity of movements is large because the human arm has many degrees of freedom and is used as a manipulandum to interact with the environment. To support a patient during the training the amount of support should be adapted in an assist-as-needed manner. We propose a method to learn the arm support needed during the training of activities of daily living (ADL) with an arm rehabilitation robot. The model learns the performance of the patient and creates an impairment space with a radial basis function network that can be used to assist the patient together with a patient-cooperative control strategy. Together with the arm robot ARMin the learning algorithm was evaluated. The results showed that the proposed model is able to learn the required arm support for different movements during ADL training. © 2011 IEEE.