Understanding how the brain learns motor skills remains a very challenging task. To elucidate the neural mechanism underlying motor learning, we assessed brain activation changes on a trial-by-trial basis during learning of a multi-joint discrete motor task (kendama task). We used multi-channel near-infrared spectroscopy (NIRS) while simultaneously measuring upper limb movement changes by using a 3D motion capture system. Fourteen right-handed participants performed the task using their right upper limb while sitting a chair. The task involved tossing a ball connected by a string to the kendama stick (picking up movement) and catching the ball in the cup attached to the stick (catching movement). Participants performed a trial every 20 s for 90 trials. We measured the hemodynamic responses [oxygenated hemoglobin (oxy-Hb) and deoxygenated hemoglobin (deoxy-Hb) signals] around the predicted location of the sensorimotor cortices on both hemispheres. Analysis of the NIRS data revealed that the magnitudes of the event-related oxy-Hb responses to each trial decreased significantly as learning progressed. Analysis of movement data revealed that integrated upper limb muscle torques decreased significantly only for the picking up movements as learning progressed, irrespective of the outcome of the trials. In contrast, dispersion of the movement patterns decreased significantly only for the catching movements in the unsuccessful trials. Furthermore, we found significant positive correlations between the changes in the magnitudes of the oxy-Hb responses and those of the integrated upper limb muscle torques during learning. Our results suggest that the decrease in cortical activation in the sensorimotor cortex reflects changes in motor commands during learning of a multi-joint discrete movement.