Break on through: Tunnel-based exploration to learn about outdoor terrain

We study the problem of designing a ground robot that learns how to navigate in natural, outdoor settings. This robot must learn, through its own experience, a traversability model. This model provides a mapping from the robot’s sensors to a traversability measure. We propose a novel approach to exploration in which the robot purposefully probes into apparent obstacles, with the goal of improving its traversability model. We formulate a decision-theoretic algorithm that identifies potential “tunnels” through apparent obstacles and estimates the benefit of exploring them. We have implemented and tested the algorithm on an outdoor robot platform and we present results from initial experiments that demonstrate the operation of our algorithm and quantify its performance advantage over a baseline system.