Writing Robot Perception And Reasoning Algorithm Research

As the important carrier of artificial intelligence,robot has played an increasingly important role.Writing robot for helping humanity write and improve writing speed and quality is of great significance,it is also an innovative and integrative application of various kinds of artificial intelligence technologies.However,from the aspect of enabling writing robots to think and learn,there are still many ubiquitous and unique difficulties existing in research,such as the multi-object segmentation algorithm and the object recognition with partial shade algorithm for helping robots to obtain knowledge about the writing scene,the reinforcement learning algorithm research to train robots learning actions,the knowledge presentation and reasoning methods for helping writing robots with commonsense reasoning,the perturbation detecting and handling methods for helping writing robots deal with perturbations,and the integration of various kinds of artificial intelligence technologies,etc.This dissertation focuses on algorithm research on knowledge acquisition,action learning,commonsense reasoning,and perturbation handling for intelligent writing robots.It aims to realize the target that writing robots could actively learn knowledge about the environment and themselves,intelligently explore the world and operate objects,understand the mission and the environment using the obtained knowledge and reasoning rules,and handle the perturbations within mission duration,so that writing robots could be more intelligent and more flexible.The main works completed by this thesis are summarized as follows:Firstly,writing robots should obtain the comprehensive visual knowledge about the writing scene,so that they could operate objects,plan missions and conduct commonsense reasoning reasonably.Considering that depth data is very beneficial for perceiving the environment and operating objects,the Kinect sensor is utilized to acquire depth images in this thesis,depth of the research mainly focuses on the depth image filtering,depth image segmentation,object size data automatic acquisition method,and the partial occluded object automatic identification method.Depth image filtering is the precondition for obtaining visual knowledge using depth images,for the problems that there are too much noise in the depth images caused by structure light and environment,we propose a combinational filter algorithm based on the pixel filtering method and the weighted moving average method,through a large number of contrast experiments we get the optimal parameters and the best filter combination style,by comparing the results of the bilateral filter algorithm and the guided filter algorithm,it is showed that our combinational filter performs better in the aspect of removing the image noise and suppressing speckle noise effect.On the foundation of depth image filter,we raise the depth image segmentation algorithm based on gradients of depth pixels and the K-mean clustering algorithm,which could segment planes with different depth and angles of multi-object in depth images after running twice K-mean clustering process.Furthermore,based on the filtered and segmented depth image and the imaging principle of Kinect,we raise the object dimension automatic acquisition method based on depth images,the acquired data is assistant with the real size of objects.To solve the problem that the object in writing scenario may be partially occluded,we propose the partially occluded object recognition algorithm based on the Bilattice reasoning framework and the combinational SVM and recognize the partially shaded chalk boxes successfully.Through comparing to results derived by the cascade of part detectors,it is proved that the method using machine learning and reasoning with priori knowledge given in our thesis has higher recognition rate.Secondly,writing robots should be able to autonomously plan and choose actions according to the writing environment,in order to train writing robots learning writing related actions based on the input visual features,the research is mainly focused on the state space representation and reinforcement learning algorithm,we propose an action study method based on the GNG(growing neural gas)network and the eNAC(episodic natural Actor-critic)algorithm.To deal with the problem that writing robots may need to turn around to face to the object of interest during the writing process,the input visual images of the writhing scene are projected into GNG network’s nodes,where different nodes correspond to different objects and denote the states of MDP problems,then eNAC,which is suitable to the MDP problems with the continuous state/action space,learn from the nodes of the GNG network,eventually the Pr2 robot used as the writing robot in Gazebo environment could find to the object of interest according to its available information.Finally,to help writing robots think about their behaviors and the mission and handling perturbations using their knowledge,in terms of writing robot common sense reasoning and perturbation handling,the research mainly focuses on detecting and handling the contradicted knowledge for commonsense reasoning and detecting & dealing with perturbations.To resolve the contradicted knowledge problem of commonsense reasoning,we propose a method to detect and dispel contradicted knowledge based on the relationship between different pieces of commonsense,which makes active logic more suitable and robust to deal with contradicted knowledge.For the perturbation problems which may probably be encountered by robots during performing tasks,we analyze the perturbation handling mechanism of MCL and realize it using the rules of active logic,the experiment about Q-learning machine enhanced by MCL performed in a reward changing world proved that MCL is sufficient and helpful for handling perturbations.Furthermore,the metacognitive integrated dual cycles(MIDCA)is incorporated as the framework for intelligent robot learning and reasoning.We raise an A-distance matrix based perturbation detection method by detecting the variation of the real value vector stream which describes the predicate state stream,the experiment about the writing miss proves that our methods are effective to detect perturbations.In the end of this part,we design two cases: the writing robot learns how far its arm is and the writing robot getting books,which uses knowledge acquisition,action learning,and reasoning all together to verify the ability of meta-cognitive reasoning and learning to help the writing robot get knowledge about itself and handle perturbations during the executing missions.