Simulation-Based Goal-Selection for Autonomous Exploration

High-level planning can be defined as the process of selection of an appropriate solution from a set of possible candidates. This process typically evaluates each candidate according to some reward function consisting of (1) cost, i.e., effort needed to accomplish the candidate and (2) the utility of accomplishing it and then selects the best one according to this evaluation. The key problem lies in the fact that the reward function can be rarely evaluated precisely. At the example of the problem of exploration of an unknown environment by a modular robot we show that precise simulation-based estimation of the cost function leads to better decisions of high-level planning and thus improves exploration process performance. Stateof-the-art techniques compute the cost function in goal-selection as a length of the path from the current robot position to a goal-candidate. This is sufficient for robots with simple kinematics for which time to reach a candidate highly correlates with a path length. As this does not hold for complex (modular) robots, we introduce the approach that generates a feasible trajectory to each goal-candidate (taking into account kinematic constrains of the robot) and determines the cost function as time needed to perform this trajectory in a simulator. The experimental results with a robot consisting of eight modules operating in several environments show that the proposed simulation-based solution outperforms standard solutions.