The primary objective of autonomous exploration is to answer the question “where to go next?”, but this question is intrinsically linked to what we are exploring and our purpose or goals. In this project, we are interested in large scale natural environments, such as parks or forests, and in the specific challenges raised when trying to explore these types of environments.

We focus on the challenges raised when trying to build an accurate 3D map of such sparse environments.
The first challenge deals with measuring the quality of a reconstruction against the ground-truth when the data is extremely sparse and where traditional methods fail. We propose a complete methodology to measure the quality of the reconstruction in both structured and unstructured environments.
The second challenge consists in predicting, or more precisely estimating, the quality of the reconstruction. More specifically, we place ourselves in the case where a ground robot needs to autonomously build a qualitative map of an unknown 3D environment, from 3D scanning LIDARs. In that case, how can we integrate the quality of the 3D-reconstruction into an exploration policy?  This is the core of this project. To do so, we prove first that the quality of the reconstruction depends on statistics on the point of view of the observations, but also that we can estimate the quality of the reconstruction from those statistics, without knowing the ground truth.

Future work will focus on integrating this prediction into an exploration policy, both in simulation and around the Symphony Lake, with our Clearpath Husky.

Work realised by Stephanie Aravecchia.