The Manta robot is a kind of underwater robot that uses only one pair of right and left pectoral fins to produce swimming propulsive forces. The measurement system of a Manta robot used in this work includes a gyro, an accelerometer and a magnetometer. This paper is devoted to discuss the attitude determination algorithm of such a Manta Robot. Since the attitude determination combining the accelerometer and magnetometer measurements is only available when the robot is in a steady-state or of uniform motion, the gyro measurement is adopted to amend the attitude in a dynamic state by an unscented Kalman filter (UKF). What's more, the gyro has a drift, so that this error will be accumulated when the time goes on. Therefore, an autoregressive (AR) model is applied to model the stochastic drift of the gyro. The output of the gyro is denoised by the Kalman filter on the basis of this drift model. Finally, some experiments with a real robot are carried out to verify this plan.