This study presents a novel torque-controlled hand exoskeleton, named HandeXos-gamma, which uses a series-elastic actuator (SEA)-based architecture to allow a compliant actuation of the hand joints, and an intention decoding algorithm that combines surface electromyography (sEMG) signals with kinematic information from the exoskeleton's encoders. The algorithm was developed offline using data acquired from healthy subjects who performed two grasping movements (lateral and power grasp) under different operating conditions while wearing the exoskeleton. Performance was evaluated for three variants of the algorithm: one using sEMG signals only, another using kinematic data only, and the last combining sEMG and kinematic data. Results indicated that the combination of the two modalities conferred greater algorithm performance than sEMG alone, thus supporting a new paradigm for adaptive robotic hand rehabilitation.