Skip to main content

Flexible, Efficient, and Scalable Autonomous Exploration and Volumetric Mapping

Time: Fri 2022-03-18 14.00

Location: U1, Brinellvägen 26, Stockholm

Language: English

Doctoral student: Daniel Duberg , Robotik, perception och lärande, RPL

Opponent: Senior Researcher Cesar Cadena, ETH Zurich

Supervisor: Patric Jensfelt, Robotik, perception och lärande, RPL

QC 20220224


Autonomous mobile robots have in recent years started to enter households in the form of autonomous vacuum cleaners and lawn mowers. The applicability of more advanced and general purpose service robots is almost endless. That is, robots that can perform a variety of tasks, instead of being specialized for a single task. To this end, there are some fundamental challenges that need to be addressed. One of the key capabilities of an autonomous mobile robot is navigation. To achieve truly autonomous navigation, the robot has to be able to localize itself, plan, execute, and update a path that takes it to its desired location, and to generate a map on-the-fly of its environment if the environment is unknown or changing. This thesis focuses on the latter two of these challenges, planning and mapping. More specifically, we investigate in the scenario where the robot lacks any prior knowledge of the environment, referred to as autonomous exploration.

One of the most important insights throughout the thesis is that these challenges should not be examined in isolation. As these are generally not the main tasks, a truly autonomous mobile robot shall perform; instead, they are necessities to fulfill higher-level tasks. Therefore, aspects such as flexibility and scalability should be regarded higher than simply accomplishing the task as efficiently or quickly as possible.

Another insight, specifically regarding mapping, comes from surveying both consumers, the ones using the maps, and producers, the ones creating the maps. Ideally, a mapping framework should be optimized towards both, as it is pointless creating maps that cannot be used as well as assuming data can be extracted from a map in ways that are unfeasible. However, in existing works this is rare. 

A third insight, specifically regarding exploration, comes from breaking down typical assumptions and simplifications that are generally applied to make the problem tractable. We show that the problem is often formulated such that it leads to unnecessary greedy behavior, where the expected information gain has too high priority. Not only do we show that with a more general formulation we can achieve better results, but also that the information gain is not important from a long-term perspective.

In this thesis, we present a mapping framework as well as an exploration framework. With these frameworks, we show that flexibility and scalability do not necessarily have to come at the cost of efficiency. We contribute the mapping framework, UFOMap, and the exploration framework, UFOExplorer, open-source to the community such that others can further develop and build upon them.