Augmented reality (AR) technology can provide immersive and natural interactive interfaces for brain-computer interface (BCI) systems. The control architecture of existing AR-BCIs is at joint-level (JL) or action-level (AL), which brings a huge user burden and poor interactive experiences. A task-level (TL) BCI control method was proposed in this study to enhance interactive experiences. The TL AR-BCI system based on steady-state visual evoked potentials (SSVEPs) was implemented controlling a robotic arm to grab and drop blocks. The online experiment of 10 subjects shows TL AR-BCI can effectively reduce the number of control steps and stimulation time while maintaining the same performance as JL and AL AR-BCIs. The performance of three AR-BCIs (JL, AL, TL) was calculated(Mean accuracy: 90.66\%, 92.52\%, and 92.2\%. Mean information transfer rates: 77.56bits/min, 80.06bis/min, and 82.71bits/min. Mean numbers of control steps: 35.48, 17.32 and 13.05. Mean stimulation time: 0.97s, 0.97s and 0.89s). The results show that TL AR-BCI can effectively reduce the number of control steps and stimulation time while maintaining the same performance as JL and AL AR-BCIs.