Learning Places in Newly Explored Environments

Abstract This paper presents a computational model of cognitive maps for navigation, which is implemented on a mobile robot. The model is based on a self-organising neural network which creates a topological map of the environment as explored by the robot. The map neurons can be seen as `place cells ' which are inspired by the discovery of real place cells in the rat's hippocampus. To include metric information, the topological map is complemented with a physical force model. The performance of the model is tested both in simulation and on the mobile robot `Samurai'. The robot is equipped with a compass as well as an omnidirectional camera, and continuously produces a low-resolution, orientation-invariant place vector. Using only this place vector and the compass information, a consistent map of a newly explored environment can be produced. 1. Introduction Robot navigation is a relatively simple task if the environment is known in advance and does not change. However, this assumption is not fulfilled in most realistic applications. Especially when robots are used as a tool to model biological navigation, whether of humans or of animals, the test environment should yield conditions similar to the real world. The simplest way to navigate only includes reactive responses. More sophisticated navigation strategies definitely include learning, and consequently memory. One method of such memory structures are cognitive or mental maps, which act as a representation of the environment that could also be coupled with actions. The term `cognitive map ' has already been introduced by Tolman in 1948 (Tolman, 1948). Electrophysiological investigations during behavioural experiments revealed some interesting correlations: In 1976, cells were detected in the rat's hippocampus which preferably fired when the rat was in a particular portion of its environment, but were largely independent of its orientation and actual view (O'Keefe, 1976). The area in the environment where one particular cell has the highest activation is called the place field of this cell. Another type of cells were found which preferably fired when th