This paper presents a prototype for a brain-controlled robot arm system using a variety of upper-limb movement imagery. To do that, we have designed the experimental environment based on brain signals. The experimental system architecture was modularized into three main components: BMI, network, and control parts. Six subjects participated in our experiments. The subject performed various upper-limb actual movement and imagery task. Each task consisted of three different movement/imagery: Arm reaching tasks, hand grasping tasks, and wrist twisting tasks. We confirmed the classification accuracies are 22.65%, 50.79%, and 54.44%, respectively. Moreover, we will demonstrate that brain-controlled robot arm system can achieve a high-level task in multi-dimensional space.