Research On Incremental Learning Method Of Developmental Robot

Developmental robot can achieve the autonomous mental development as the human infant,and just needs to edit a certain set of the developmental programming for obtaining the variety of skills.Based on the autonomous developmental algorithm,the robot can acquire the abilities to perform various kinds of unknown tasks through the continuous learning.It can solve the problems of the programming for the non-specific task,multiple function,and strong adaptability and so on.Therefore,Developmental robot with the autonomous learning has become a research hotspot and attracted more and more attention of researchers.In this paper,the algorithms of the visual feature extraction and the autonomous development were proposed and verified by experiments for the developmental robot.Firstly,a new visual feature extraction algorithm is proposed.The developmental robot is needed to incrementally learn in real-time for autonomously accumulating the feature.The robot can obtain the outside information and then implement the image processing from its visual system.However,the camera of the robot has a high resolution in general,and the dimension of the sequent arriving frame-images is large from the video system.The robot can not implement the image processing in real-time for its limited computation abilities.Therefore,it is necessary to extract the image features,to decrease the image dimensions,and to reduce the computation time.In this paper,to extract the image feature,a novel incremental bi-directional principal component analysis algorithm is proposed based on the existing algorithms of the candid covariance-free incremental principal component analysis and the bi-directional principal component analysis.The proposed algorithm has not the recursive computation ability,but also the incremental computational ability,for approximately estimating the two-dimensional image matrix in direct.It can effectively reduce the computational amount and greatly shorten the running time for the whole feature exaction,and meet the requirements of the real-time computation for the video stream of the developmental robot.The key of the autonomous learning for the developmental robot is the autonomous development algorithm.Based on the autonomous developmental neural network,an autonomous development learning algorithm is presented to simulate the working mechanism of the brain visual cortex for human being.The lobe component analysis algorithm is adopted to implement the autonomous development process.The pre-response value is used to balance and match between the input of the network and the immediate knowledge expression of the network.The top-k competition learning is also used to simulate the neuronal inhibitory effect of the brain.And the Hebbian learning can be used to update the weight vectors of the neurons that successfully compete with each other from the top-k method.Considering the amnesic time,the neuronal ages is introduced to the developmental neural network.Based on the proposed autonomous developmental neural network,the developmental robot can learn by inputting the sample images,what it has learned will be stored into the neural network and form a stable memory.Taken the six degree of freedom reconfigurable manipulator as the experimental platform,the grasped blocks were used to verify the cognitive learning of the robot.The experiments is based on the previous studies of the visual feature extraction and the autonomous development network.To verify the effectiveness of the proposed algorithms,the sample images of these blocks were cropped from the videos of the robot camera.Two kinds of experiments were implemented.First,in the feature extraction experiments,the sample images were input gradually into the algorithms by the incremental form.The proposed algorithm and other existing two algorithms were performed to compare the performances of these algorithms.The experimental results show that the classification rate of the proposed incremental bidirectional principal component analysis algorithm can reach more than 90%,higher than those of the other two algorithms.Simultaneously,the computing time of the proposed one is far smaller than those of the other two.The proposed one can meet the requirements of the real-time computation.Second,in the experiments of the autonomous developmental neural network,the standard handwritten digital data set and the sample images of the grasped blocks were used as the experimental objects,respectively.The visualizations of the weight vectors of neurons of the two experiments were also shown subsequently.The classification and recognition experiments were performed and realized.The proposed autonomous developmental neural network can incrementally learn and recognize for the developmental robot.