Internal Predictive Model In Cerebellum For Balance Control Of Robotic Agents

Cognition plays a key role in acquiring motor skill including motor control and motor learning. To understand and simulate the cognitive behaviors of biological systems and try to endue these to robotic agents is the aim of this thesis. The first most important factor in autonomous robotic agents control is human posture balance control. The research of a computational cerebellar model is essential to robotic agents to simulate human posture balance control.In order to cope with the balance control problem of robotic agents, this thesis proposed a kind of internal predictive model (IPM), which can be represented by a cerebellar model based on Kalman estimator and a cerebellar model based on Smith predictor, according to the neurobiological evidences with theories of Kalman filter and Simth predictor in the framework of internal model, respectively. The main research works of this thesis can be summarized as follows:1) A model of human posture balanceThrough an analysis of human posture balance control, a model of simplified model of human balance is introduced, and then simplified it in the form of inverted pendulum. By comparing the experimental results with proportional-derivative (PD) controller and neural network proportional-derivative (NNPD) controller to the balance control of cart-pole system, we draw a conclusion: human balance control needs a kind of self-tuning adaptive IPM in cerebellum during the process of motor learning to achieve the effect of motor control, to help robotic agents acquire the motor skills as body balance.2) A cerebellar model based on Kalman estimatorWe proposed a cerebellar model with a Kalman estimator based on the conception of IPM, which can ensure the accuracy and stability of the system during the motor control processing when it was applied to the body balance problem. This forward model is performed by a neural network which is supervised by the teacher signals from the estimator and feedback model, while the feedback model employs the PD controller to ensure the overall stability of the system. The effectiveness of this model on the balance control of robotic agents can be illuminated by the simulation and experimental results with the inverted pendulum.3) A cerebellar model based on Smith predictorA cerebellar model based on Smith predictor as another kind IPM is proposed in this thesis, which aims to solve the problems caused by the long loop delays in a feedback control system with a high gain from a physiological view. The mechanical and neural feedback gains must be quite high for an effective control effect. Meanwhile the transported delay from those biological processes should be noticed. Thus, this proposed model is reasonable in conception. The two forward models in this scheme could be acquired from neural learning. The training signal inspired by recent cerebellum research is used to learn those dynamic models, and the experiments with inverted pendulum testify the good performance of this idea on balance control of Robotic agents. Because the predictive models are required to represent controlled object and time delay accurately, this scheme is sensitive to parameters. The inverted pendulum is a kind of typical nonlinear and unstable system, so some exploration research works are presented.Simulation experimental results in this thesis demonstrate that internal predictive models can assume the balance control task of robotic agents. These results prove that the proposed models can be understood and explained with the real neurobiological systems functionally.This research is supported by National Natural Science Foundation of China (60375017), Funding Project for Academic Human Resources Development in Institutions of Higher Learning under the Jurisdiction of Beijing Municipality (PHR (IHLB)), and the Special Founding for Doctoral Program (20050005002). Related work has been approved for publication by the publication of Lecture Notes in Computer Science (LNCS, SCI indexed), the Proceeding of 6th World Congress on Intelligent Control and Automation (EI indexed), and the Journal of Tsinghua University (Nature Science and Technology) (EI source). During the time of graduate research, I attend the third International Symposium on Neural Networks (ISNN) and gave an oral representation. The study presented in this thesis may make a significant contribution to the application of motor learning model to control system, and could be applied to the field of machine learning, automation and robotics.