Research On Intrinsic Motivation System Of Developmental Robot

The movement balance control abilities of the biology are derived from the study of sensorimotor nervous system, and intrinsic motivation system plays an important role. Exploring the intrinsic motivation mechanism in the sensorimotor nervous system, formalizing the intrinsic motivation theory, simulating and copying the structure and function of the self-organizing control mechanism, applying it to the robot system, through learning and training of the robot, gradually obtaining the autonomous learning capacities and movement balance control skills similar to humans and animals, which are the main topics of the current robotics, cognitive science, psychology, neurophysiology, and control science.The combines the cognitive science, psychology and neurophysiology to establish the intrinsic motivation models with different network structure from the angle of the bionic, and explore the autonomous learning and control mechanisms of the sensorimotor neural systems. Using the two-wheels of self-balancing robot experiment platform, the thesis studies the autonomous development and learning abilities of the bionic model based on the intrinsic motivation in the process of the movement balance control of robot. The main research achievements of the thesis are as follows:1) Intrinsic motivation model based on the operate conditioningThis thesis puts forward a kind of intrinsic motivation bionic model based on the operate conditioning principle (OC-IM). Under the theory framework of the operate conditioning, the intrinsic drive capabilities about the autonomous development and learning of the biological sensorimotor nervous system are embodied by the form of mathematical model of the neural computation, and applied in the two-wheeled robot. Through the autonomous development and learning, the robot can choose the optimal behavior probably, and maintain the movement balance state. The simulation results show that the intrinsic motivation bionic model can make the two-wheeled robot successfully obtain the balance control abilities through learning and training, which shows the good learning abilities about autonomously controlling the movement balance.2) Intrinsic motivation bionic model with memory growthThis thesis proposes the intrinsic motivation bionic model with memory growth (GCS-IM). The model can insert the new neurons to the competitive layer according to the accumulation of experiences and the needs of task requirements, increase the cell numbers of the network, and adjust the size and structure of the network autonomously and dynamically. According to the physiological behavior about the action selection probably in the process of cognitive development of the biology, the model uses the Boltzmann machine to optimize probably the action option, highlights the random choice behavior autonomous process of the intrinsic motivation, and applies it to the movement balance control of the robot. The simulation results show that the algorithm can achieve the desired control objectives, and the learning speed and dynamic performance show the better control performances, which reflect the cognitive process similar to the intrinsic motivation of the human and animal.3) Intrinsic motivation bionic model with memory pruningDue to the complexity of the different tasks, the structure of the neural network is difficult to determine. Sometimes it is too complicated so as to lead to the lower efficiency and generalization ability, and easy to appear over fitting phenomenon. This thesis designs a kind of intrinsic motivation bionic model with memory pruning (OBS-IM), and it contains the evaluation neural network and the action neural network. Aiming at the problems of the large scale neural networks, this model can delete the multiple neuron weights at the same time, adjust the remaining weights of neuron, achieve the rapid network structure adjustment speed by changing the Hessian matrix method to adjust the network error function. Applying the model in the movement balance control of the two-wheeled robot, the simulation experimental results show that the intrinsic motivation bionic model with memory pruning (OBS-IM) has the faster learning rate and the better dynamic performance, and at the same time, embodies the cognitive and learning process similar to the biological intrinsic motivation.4) Sensorimotor system of the development robot based on the intrinsic motivationBecause of the complexity of the sensorimotor system, the thesis proposes a sensorimotor system based on the intrinsic motivation under the framework of the growth or pruning dynamic structure neural network. This model can determine increasing or reducing the number of neurons according to the size of the neural node sensitivity values about the hidden layer, improve the network learning rate and enhance the ability of dealing with complex networks. The cognitive model is mainly composed of three modules:action control module, prediction module and buffer module. Among them, the action control module can realize the mapping from the states to the actions, the prediction module is for the probability selection of the actions, and the buffer module is to receive and send the useful signal to the action control module and the prediction module, so as to make the function of the sensorimotor system realize. Applying the sensorimotor system based on the intrinsic motivation to the balance control of the two-wheeled robot, the simulation experimental results show that the cognitive model can improve the movement balance control skills by autonomous learning and development unceasingly, and verify that the model has the strong ability of the cognition and development.