Contrary to existing paradigms describing robot motion which consider a small number of missions executed in a finite environment, LEN considers the environment in which the robot evolves as potentially inifinite. Moreover, the robot is never reset, similarly to living beings. It never stops learning and forgetting parts of the environment. LEN requires an integration of SLAM, path planning and physical navigation (which we call PNSLAM), where integration means that none ot these three problems can be considered independently of the others. For instance, the SLAM algorithm may request a trajectory change to the path planning algorithm in order to confirm that the robot is where it thinks it is. Figure 1 below shows how LEN can be built from existing approaches.

Figure 1: Lifelong Exploratory Navigation - general architecture.

Implementation of LEN on a mobile robot navigating in an essentially 2D world is described in Figure 2. This implementation was tested on a Pioneer 3 robot (Figure 3).

Figure 2: Lifelong Exploratory Navigation - 2D implementation. Space scales in red, execution time in blue and algorithmic complexities in green, with N the size of the environment (or number of observed features). I heavily modified or coded from scratch the blocks marked with a yellow symbol.

Figure 3: The pioneer 3 robot equipped with a Kinect 1 camera (used only for depth sensing) and a simple MEMS magnetometer mounted on top of a cardboard box (to avoid magnetic interference from other components aboard the robot) was used for Lifelong Exploratory Navigation experiments.

(less)