Several Researches About Intelligent Robot With Perception And Cognition

The research on intelligent robot relates to several different areas such as computer science, information science, artificial intelligence, mechanical engineering and control theory etc. Although many robots have been applied in industrial fields, compared with human brain, their intelligence is very low. Many problems, that are easy for human, are still a great challenge for existing intelligent robot technology.This paper focuses on several hot problems in the studies of intelligent robot, especially on some up-to-date works about developmental perception, cognition and human brain simulation. Our work includes: 1) autonomous navigation, self -localization and digital map building for the mobile robot with Laser; 2) the developing model of cognitive mapping for the intelligent robot; 3) the biologically plausible computational model of sensory mapping for intelligent robot. Based on the traditional methods, we present some new algorithms about autonomous navigation, self-localization and digital map building. Besides, some new methods, such as Autonomous Mental Development (AMD) and the simulation of biological neural system, are used to improve the perception and cognition of the intelligent robot and to construct some creative cognitive and sensory models. These works are organized as follows.In the first part, we take two problems: 1) The inherent limitations of Artificial Potential Field method (APF) for the autonomous navigation in unknown environments. For this problem, we propose a new concept—"Behavior Information Potential Field". Based on it, a punishment functions are presented to implement path planning in unknown environments. Our method breaks some limitations of the classical navigation methods APF, such as the "trap situation due to local minima" and "no passages between closely spaced obstacles". 2) The challenges of precision and computation speed in "Simultaneous Localization and Map-building" (SLAM). To solve this problem, we present new algorithms of Coarse-to-fine classification and vertex localization to improve the traditional SLAM methods.In the second part, we focus on the lower regression rate and irrationality of the Hierarchical Discriminate Regression algorithm (HDR) which is based on AMD. 1) We proposed a Fisher-HDR classifying neural network, which combined Linear Discriminate Analysis (LDA) and HDR together and increase the regression rate of HDR. 2) A new strategy for incremental LDA is also proposed by us to enable Fisher-HDR neural networks to process the data stream adaptively. Our method is based on two-PCA structure and natural power method of adaptive PCA. Compared with other existing incremental LDA algorithms, our method is better in the convergent speed and time-consuming.The final part is about the sensory mapping of intelligent robot. Now, a new trend in artificial intelligence school is the simulations of human brain, which provides us a promising approach to make robots more intelligent. Based on it, we follow the AMD method and build the sensory mapping by modeling retina and primary visual cortex. 1) We proposed a multi-layer neural networks model with temporary development for human retina. Compared with other existing models, our model is more similar to retina in the layered structure and the information processing mode. Our model can change adaptively the local contract, sensitivity and sharpness of vision and also improve adaptability of this model by simulating the retinal predictive coding that discovered recently by Hosoya in the journal "Nature". 2) We present a spatiotemporal developmental neuronal cluster model to mimic the feature extraction and neuronal development in primary visual cortex. This model adapts the synapses of both spatial receptive fields (RF) and temporal RF by a unifying unsupervised learning rule. This is the first work that we know that models the temporal RF's characteristics of networked visual neurons. It can simulate the development of orientation selective neurons and chromatic antagonistic neurons in primary visual cortex. The developmental visual neurons have a good performance in feature extraction. 3) We proposed a hypothetic model with lateral inhibitions and give corresponding mathematical deduction to illustrate the internal biological mechanism of neuronal temporal RF. 4) Using the proposed neuronal cluster, we construct a multi-layer bottom-up sensory network (MBSN) to realize the visual attention selection. The top-down inhibition passage based on the novelty of stimuli is added to the MBSN to drive the focus of attention. Our model extracts three kinds of features—color, motion and edges—by a group of developmental neurons that adapts to the environments. Our model can deal with simultaneously the sequence combined of still images and dynamic videos. Moreover, the networked structure makes it easier to be implemented by hardware. This model is helpful for saving computational resource and improving the visual perception of intelligent robots.