Balancing the Information Gain Against the Movement Cost for Multi-robot Frontier Exploration

Abstract

This article investigates the scenario where a small team of robots needs to explore a hypothetical disaster site. The challenge faced by the robot-team is to coordinate their actions such that they efficiently explore the environment in their search for victims. A popular paradigm for the exploration problem is based on the notion of frontiers: the boundaries of the current map from where robots can enter yet unexplored area. Coordinating multiple robots is then about intelligently assigning frontiers to robots. Typically, the assignment of a particular frontier to a particular robot is governed by a cost measure, e.g. the movement costs for the robot to reach the frontier. In more recent approaches these costs are traded off with the potential gain in information if the frontier would be explored by the robot. In this paper we will further investigate the effect of balancing movement costs with information gains while assigning frontiers to robots. In our experiments we will illustrate how various choices for this balance can have a significant impact on the exploratory behavior exposed by the robot team.

3-D exploration with an air-ground robotic system

“…This approach has been expanded to cover multi-robot teams [2], [3], [4], [5], [6] with great success and this work is an extension of our earlier work in this category [7], [8]. However, most of these approaches, including our previous work, only deal with 2-dimensional (planar) exploration.…”

“…CountM ap 4 The inputs to the GETGOAL function are the x, y, z, θ poses p of all of the robots and a parameter list r providing the footprint, nominal altitude (= 0 for UGV), sensor field of view, and sensor position and orientation relative to the robot body frame, for each robot.…”

“…Once these two initial processing steps are completed, the map is parsed to determine all of the frontier cells. A standard definition of a frontier cell is an unknown cell that is directly adjacent 4 to a known free cell. We use a slightly modified definition of this in that we consider a visually cleared cell as known, and all others, even if we have laser data on that cell, as unknown (line 6).…”

Planning for multi-robot exploration with multiple objective utility functions

“…The non-obstacle portions of this boundary are collectively known as the frontier [1] and are used as candidate goal points. Algorithms have been proposed that extend the basic frontier exploration scheme to multi-robot teams [2], [3], [4], [5] and outdoor environments [6].…”

“…It must be recognized that as each state is entered it may change the amount of certainty in surrounding states and thus lower the information gain for subsequent states along the trajectory. Due to this non-Markov property optimizing the information gain over trajectories is computationally expensive, and many algorithms estimate the information gain of a trajectory by using the information gain of the final point [4], [5], [7] while a few have attempted to explicitly include the expected information gain along the entire trajectory [3], [8]. Others have combined the information gain with additional features such as communications constraints [9] or improved physical models such as the likelihood of specular reflection from nearby obstacles [10].…”