Restoring lost functionality and overcoming the disability resulting after stroke is an ongoing challenge in the field of rehabilitation. The most common disability is hemiparesis, i.e., weakness in one half of the body contralateral to the side of the stroke. Although multiple forms of therapy exist there is no clear understanding of the principle governing recovery. The proposed work aims to fill the gaps in motor recovery and provide insights into how to design an optimal rehabilitation protocol for a given patient. The first study in this work consists of a bimanual convolutional neural network to which two laterally shifted images of a target in 3D space were given as input. Muscle activations required to reach the target were obtained as output from the network. To this network, lesion was incorporated, and rehabilitation was provided. The recovery achieved under different rehabilitation paradigms was used to map parameters of rehabilitation to characteristics of lesion. In the second study, the aim was to understand the division of motor control between the 2 hemispheres of the brain. Although contralateral control is often assumed, there is considerable bilateral activity found even during unimanual movement. Results from 2 experimental studies where the subjects undergo adaptation to visuomotor rotation were used for replication in this study. The study used multiple feedforward recurrent networks and showed that the different motor variables are handled differently by both the hemispheres – distance to be moved encoded by the non-dominant, and direction by the dominant hemisphere. In the third study, reinforcement learning algorithm – q-learning, is used to develop a computational agent capable of playing a virtual tennis game. The same game was also deployed to stroke patients, and their behaviour was recorded. The learning achieved by the agent will then be compared to that achieved by the patients.