The initiative of two-wheeled self-balancing robot in the process of learning is poor. Inspired by intrinsic motivation theory of psychology, an autonomous development algorithm for intelligent robot based on intrinsic motivation was put forward. In the frame work of the reinforcement learning theory, the algorithm introduced human curiosity of intrinsic motivation theory as the internal driving force, and external reward signal into entire learning progress, and adopted double internal regression neural network for storage of knowledge learning and accumulation, which made robot gradually learn autonomous balance skill. Finally, aiming at the effects of measurement noise pollution on two-wheeled angular velocity of robot, further by adopting the method of Kalman filter to compensate, to speed up the algorithm convergence, and reduce the system error. Simulation experiments show that this algorithm can make the two-wheeled robot obtain cognition through interaction with the environment, therefore successfully learn balance control skill.