Robotic-assisted rehabilitation of the upper limb following neurological injury can achieve best possible functional recovery when patients are engaged in the therapy. However, implementation of active training is still difficult as it’s challenging to detect human motion intention online and impose corresponding robot control. This paper introduces a novel upper-limb rehabilitation robot, and proposes a sEMGdriven (sEMG: surface Electromyography) torque estimation model based on artificial neural networks (ANN). The robot has three DOFs, of which the first two DOFs adopt a planar parallel structure, and the wrist module has an exoskeleton form. In this study, we design an impedance controller and an admittance controller for the first two DOFs and the wrist module, respectively. Specifically, for the first two DOFs, the assistance/resistance force at the end-effector was controlled according to its motions and desired interaction impedance; for the wrist module, an sEMG armband was used to collect 8 channels of sEMG signals from the forearm muscles, and a time-delay ANN model was developed to estimate the wrist pronation/supination torque, based on which the wrist rotation was controlled according to the human motion intention. To overcome the overfitting problem, besides the experimental samples of wrist rotation, both resting and co-contraction samples were collected for training. Finally, combining with the design of a virtual reality game and force fields, the proposed methods were implemented and tested experimentally on the upper-limb rehabilitation robot.