Learning And Control Method For Robotic Cognitive Behavior Based On Episodic Memory

Based on the ideology and methods of behaviourism, this dissertation focuses on the two main issues of robotics from the perspective of episodic memory in cognitive neuroscience, which are behavior sequences learning and behavior selection. Currently, robotics research for specific tasks under structured environment has many research results. But the increasing demand requires robot performing non-specific tasks under complex dynamic environment with uncertainty. With the requirments of human-level intelligence, cognitive robot, a new field of research, comes into being. This thesis proposes models and methods for some scientific problems in cognitive robot, such as robot generalized environmental modeling, real-time online learning and real-time optimized behavior, by utilizing episodic memory in biology and cognitive neuroscience.Considering the biological basis of episodic memory-related hippocampal neurons activation, the mathematical model of robot episodic memory for generalized environmental modeling is tried to establish. For uncertain environment, mapping for multi-dimensional perception to one-dimensional state neurons is built, and the framework of episodic memory-driving Markov decision process (EM-MDP) integrating neuronal excitation mechanism is proposed. Also the cognitive behavioral control architecture is proposed based on visual attention and episodic memory, and the implementation process of proactive cognitive behavior and characteristics of this architecture are analyzed.Aiming at the problem of real-time storage, incremental accumulation and integration for robotic experience and knowledge, a generalized incremental learning and real-time online learning method based on episodic memory is proposed to simulate the organization of episodic memory. Biology-inspired attention mechanism is utilized to obtain stable natural landmarks of input scenes which dynamic information is eliminated, and local binary pattern (LBP) is utilized to acquire feature sequences of landmarks. Inspired by adaptive resonance theory (ART) and sparsely distributed memory (SDM), the process of autonomous online learning for episodic memory network is presented by Hebbian rule, and uncertain information of the system is processed. Then the incremental task-specified episodic memory network is built. Experimental results of cognitive learning for robotic environment based on EM-MDP framework shows that the algorithm has robustness to uncertainty and can realize memory accumulation, integration and updating.Aiming at the problem of whole target selective attention and robotic applications under redundant environmental information, a target-driving object-based visual attention servo control method is proposed. Descriptive model of task target is built, and priori knowledge is obtained using GMM feature deduction. Through the introduction of proto-object and bias feature template, a bias attention model based on object is proposed for searching and selective attention of the whole task objective. The saliency map is obtained based on bias attention, and a visual attention servo control method is developed by extending image-based visual servo (IBVS) algorithm to the field of robotic cognitive control, to realize the tracking and closing for potential target for further processed recognition or manipulation. Experimental results show that the algorithm is appropriated for task target-directd robotic applications.A cognitive behavioral planning and control method based on episodic memory is proposed for non-specific tasks under uncertain environment. Adaptive behavior planning, prediction and reasoning are achieved between tasks, environment and threats. Firstly, the algorithm for robotic global planning and behavior prediction based on EM-MDP is developed utilizing neuron synaptic potential. By this algorithm, robot can evaluate the past events sequence, predict the current state and plan the desired behavior, avoiding the problem of curse of dimensionality and perceptual aliasing in partially observable Markov decision processes (POMDP). Secondly, a local behavioral planning approach based on risk function and feasible paths is employed to achieve path optimization and behavior reasoning under conditions of imperfect memory. At the meantime, a cognition inspired navigation algorithm for memorial path correction is proposed by scale invariant feature transform (SIFT) feature of salient landmrks, to make up for the deficiency of event localization only through LBP feature. Thus, by the premise of real-time performance, the robustness of features can be ensured by further improving feature dimensions of landmarks. Finally, real-time optimized behavior is generated by the cognitive behavioral decision-making strategy. The effectiveness of the proposed methods is proved by simulating complex tasks in actual general scenes.