Creating a humanoid intelligent robot is the ultimate goal of Robotics and Artificial Intelligence. Developmental robotics which based on autonomous mental development has become a novel paradigm and provided a new way to the goal. Developmental robot simulates the developmental process of human, which makes the programmers get rid of the task-oriented programming and keep free from the heavy physical fatigue and mental exertion. Meanwhile, developmental robot has the ability of online learning. The series of advantages promote it has been a hotspot in the robotics field. However, task-conflict and the generalization ability as two common disadvantages exit in nearly all of the existing developmental models. How to solve the problem becomes the top priority for researchers.The dissertation firstly introduces the concept of developmental robot, summarizes the current research advance home and abroad, and emphasizes the main study methods and disadvantages. Ensemble learning is a novel machine learning paradigm, which could significantly improve the generalization ability of learning systems by utilizing multiple learners to solve a problem. So ensemble learning is used in the developmental algorithm in the dissertation. Furthermore, in order to solve the problems of ask-conflict and the generalization ability, TDD (Task-driven Developmental Algorithm) and NCETDD (Negative Correlation Ensemble Task-driven Developmental Algorithm) algorithms are proposed. NCETDD integrates BP neural network with negative correlation learning based on TDD. The experiments test TDD in multi-task, and compare it with IHDR in two aspects: real time and task conflict. NCETDD is tested not only in above three aspects but also in the generalization ability. Experiment results show that NCETDD algorithm can solve the problem of task confliction effectively and improve the generalization ability of developmental robot. Last, some future works are given. |