Development of a cognitive map for mobile robot and its advancement inspired by place cells in hippocampus

1.通过适当设置物体的仰角，计算出曲面的倾角。将得到的角度作为自组织映射的输入，我们训练了一维SOM。迭代计算提供了一种全向反射镜的设计，其分辨率与仰角的概率密度函数成比例。还发现，在某些概率密度函数中，使用凸面镜不可能实现与概率密度函数成比例的分辨率，但使用凹凸镜是可能的.利用全景图像的局部高阶自相关函数的平移不变性等价于全向图像的极坐标局部高阶自相关函数（PHLAC）的旋转不变性，成功地获得了旋转不变性特征。将35维PHLAC特征与粒子滤波相结合，实现了对移动的机器人Khepera II的实时定位（位置和姿态），平均位置误差为31 mm，姿态误差为5.5 °.我们成功地分割使用模块化网络自组织映射（mnSOM），并在新兴的三个抽象概念，即，直线运动、右转弯和左转弯。应用新的数据产生的mnSOM实现了95.2%的正确分割率。基于所得到的mnSOM，我们还成功地构建了一个简单的环境，从基于网格的地图基于图的地图。在整合不同位置的感觉信息时，当障碍物逐渐移动时会发生数据失配。我们发现，前向传感器模型和EM算法的组合可以解决数据不匹配。

1. Appropriately setting an angle of elevation of an object, we calculate the angle of inclination of curved surface. Providing the resulting angle as an input to a self-organizing map, we train a one-dimensional SOM. Iterative computation provides a design of an omni-directional mirror with resolution proportional to probability density function of an angle of elevation. It also turned out that, in some probability density functions, realization of the resolution proportional to probability density function is not possible using a convex mirror, but is possible using a convex and concave mirror.2. Taking advantage of the fact that translation invariance of local higher-order autocorrelation functions for panoramic images is equivalent to rotation invariance of polar local higher-order autocorrelation functions(PHLAC) for omni-directional images, we succeeded in obtaining rotation invariant features. Combination of these 35-dimensional PHLAC features and particle filters realized real-time localization (position and orientation) of a mobile robot, Khepera II, with average position error of 31mm and orientation error of 5.5 degrees.3. We succeeded in segmentation using modular network self-organizing maps (mnSOM), and in emerging three abstract concepts, i.e., straight movement, right turn and left turn, based on sensory data. Application of novel data to the resulting mnSOM realized the correct segmentation rate of 95.2%. Based on the resulting mnSOM, we also succeeded in constructing a graph-based map from a grid-based map for a simple environment.4. In integrating sensory information at different locations, data mismatch occurs when an obstacle gradually moves. We found that combination of forward sensory models and an EM algorithm can resolve the data mismatch.