An Evolutionary Navigator for Autonomous Agents on Unknown Large-Scale Environments

— Computation of a collision­free path for a movable object among obstacles is an important problem in the fields of robotics. In previous research we have introduced an evolutionary algorithm for a robot moving on a known map considering a 4­connected grid model, and we have obtained encouraging results. In this paper, we focus our attention on a more complex motion planning problem: An autonomous agent with a limited sensor capability which is moving in a completely unknown large­scale environment. We introduce an evolutionary approach that has shown some adaptation abilities due to its constant update of its environment knowledge, and re­planning only when it is strictly required. We compare our approach for various map sizes to a very well­known evolutionary algorithm and to the complete approach D* Lite. Our algorithm outperforms them in both CPU time and in the number of re­plannings. 1INTRODUCTION Computation of a collision­free path for a movable object among obstacles is an important problem in the fields of robotics. The motion planning problem in known environments, or off­line planning, has been thoroughly studied [13, 15, 17]. Nowadays, there are many approaches that can successfully solve various instances of this problem. Complete methods such as A* and its variations are the most useful in this case [13, 14]. The biggest advantage of A* is that it strongly uses the knowledge of the object's position on the map. This knowledge allows it to quickly reduce the search space by discarding the unavailable branches of the search tree. It was designed to solve the static version of the motion planning problem. Inspired by A*, Stentz et al. [5, 11] and Koenig [6, 8] have proposed Dynamic A*, named with the acronym D*. It has been designed to solve motion planning problems in unknown environments, named on­line planning. It re­plans each time the robot is in front of a previously unknown object. Given the new robot sensorial information the algorithm re­plans its collision­free path by cutting the search tree in the same way as A* does. In this paper, we are interested in tackling the robot navigation problem with the following two important characteristics: • The robot is positioned in an unknown and large­scale environment. • The robot has a short sensorial capability with respect to its environment size, with a sensorial radius of just two cells around it.