Understanding Greediness in Map-Predictive Exploration Planning

Robot Exploration Using Knowledge of Inaccurate Floor Plans

“…Indeed, the later stages of a typical exploration process often provide little contribution to the completion of the final map, at a high cost. As noted in [5], in some settings, up to 71% of the total exploration time is spent 1 M. Luperto is with the Dipartimento di Informatica of the Università degli Studi di Milano, Italy matteo.luperto@unimi.it 2 M. Ferrara, G. Boracchi, and F. Amigoni are with the Dipartimento di Elettronica, Informazione e Bioingegneria of the Politecnico di Milano, Italy marcomaria.ferrara@mail.polimi.it, {giacomo.boracchi,francesco.amigoni}@polimi.it Fig. 1: The map after t = 90 min of exploration (a) already represents almost all of the environment (the portion of the mapped area is A t = 0.98), except a few uninteresting corners and portions of rooms scattered across the whole environment, highlighted in blue in (a).…”

“…Other works end the exploration when no candidate location is left [14]. A positive aspect of this stopping criterion is that it guarantees that the map is complete; on the negative side, this completeness is obtained at the expense of a long exploration time, where most of the time is spent reaching far-away uninteresting locations that have been left behind in the initial part of exploration [5], [15]. This issue is often the result of fast exploration strategies that prioritize reaching frontiers with large information gains and, consequently, encourage aggressive choices to quickly explore unobserved parts of the environments [15]- [18].…”

“…This issue is often the result of fast exploration strategies that prioritize reaching frontiers with large information gains and, consequently, encourage aggressive choices to quickly explore unobserved parts of the environments [15]- [18]. As pointed out in [5], these strategies allow a robot to make greedy choices to map most of the environment in a short time but, from that moment on, the robot can spend up to 70% of the total time in exploring marginal details of the environment that have been left behind, but that constitute about 10% of the area of the environment. An example is shown in Fig.…”

FloorGenT: Generative Vector Graphic Model of Floor Plans for Robotics

“…Autonomous exploration planning is an obvious example of a classical robotics problem where such a prediction model should be immediately applicable; however, we have previously shown that traditional non-predictive exploration planners are not well-suited to using predictions, and predictions can actually have a negative impact on exploration performance [5]. In this article, we have limited the scope to evaluating predictions by the primary variable that they affect in the exploration context: predicted information gain.…”