Information Gain Is Not All You Need

TL;DR

This paper proposes a novel heuristic, distance advantage, which selects candidate frontiers based on a trade-off between proximity to the robot and remoteness from other frontiers, and shows that distance advantage significantly reduces total path length across a variety of environments, both with and without access to prior map predictions.

Abstract

Autonomous exploration in mobile robotics often involves a trade-off between two objectives: maximizing environmental coverage and minimizing the total path length. In the widely used information gain paradigm, exploration is guided by the expected value of observations. While this approach is effective under budget-constrained settings--where only a limited number of observations can be made--it fails to align with quality-constrained scenarios, in which the robot must fully explore the environment to a desired level of certainty or quality. In such cases, total information gain is effectively fixed, and maximizing it per step can lead to inefficient, greedy behavior and unnecessary backtracking. This paper argues that information gain should not serve as an optimization objective in quality-constrained exploration. Instead, it should be used to filter viable candidate actions. We propose a novel heuristic, distance advantage, which selects candidate frontiers based on a trade-off between proximity to the robot and remoteness from other frontiers. This heuristic aims to reduce future detours by prioritizing exploration of isolated regions before the robot's opportunity to visit them efficiently has passed. We evaluate our method in simulated environments against classical frontier-based exploration and gain-maximizing approaches. Results show that distance advantage significantly reduces total path length across a variety of environments, both with and without access to prior map predictions. Our findings challenge the assumption that more accurate gain estimation improves performance and offer a more suitable alternative for the quality-constrained exploration paradigm.

Figures (6)

• Figure 1: Distance at completion $d_T$ for a selection of gain affinities $\lambda$, where a higher $\lambda$ means stronger preference for gain and a lesser concern with the length of the path to acquire it. Naive gain refers to the assumption that unknown space is occlusion-free, i.e., yields maximal gain; in true gain, the real would-be sensor scan is used for gain computation. Tellingly, negative affinities, i.e., minimizing gain, results in a lower $d_T$ than maximization, and no choice is substantially better than nearest frontier, i.e., $\lambda=0$.
• Figure 2: Illustration of distance advantage in the beginning of exploration. The robot (star) preferentially explores frontiers (solid coloring) with higher distance advantage. It is heading towards a closed off room because it is nearer that region than it would be from most other places. By contrast, its distance to the corridor is higher than it would be elsewhere, repelling it from that region.
• Figure 3: Data is collected in three diverse environments: a large office from a real-world floor plan with both small cubicles and large lecture halls, a non-rectilinear cave environment with many small pockets, and a labyrinth-like maze with both shallow and deep dead-ends. Pink circles indicate starting locations, the light blue region depicts a sensor scan from the point of view of an example starting location indicated by the brown star polygon. The zoomed region is the same size as a local window for the planner.
• Figure 4: Comparison of coverage $c(d)$ and total frontier size $f(d)$ as functions of distance traveled $d$. The shaded areas indicates an 80 confidence interval, the solid line indicates the mean. Data collected across $10$ runs for each method from different starting locations in the office environment.
• Figure 5: The effect of prediction range $c_p$ on completion distance $d_T$ in the office environment. Data collected across 10 runs from different starting locations, for each method and prediction range. Error bars represent one standard deviation.