IEEE Robotics & Automation Magazine - September 2020 - 109

devices and HAZMAT suits, which can be time-consuming
and places personnel at risk.
The optimal response to releases of HAZMAT would be
standoff, rapid, and reliable, thus keeping personnel out of
danger while maximizing the efficiency of the response. We
propose an autonomous unmanned aerial vehicle (UAV)
with integrated HAZMAT sensors as a system providing a
reconnaissance and survey capability for obtaining real-time,
targeted (i.e., most informative) HAZMAT measurements
without endangering personnel. This article presents our
recent research and experiments undertaken to develop such
a system.
It is most common to perform source term estimation
(STE) using a network of static detectors on the ground [1].
However, the focus of this article is on estimating the source
term of a release using a UAV equipped with an in situ atmospheric concentration sensor. An unmanned aerial system
(UAS) can supplement an existing network, or replace it
entirely, by intelligently gathering spatiotemporally indexed
concentration measurements of the release. There is limited
research in the area of STE using a UAS. However, it has
recently been validated in an experimental study previously
conducted by the authors [2]. In this study, a systematic flight
pattern was executed to gather the HAZMAT data used in a
Bayesian inference-based estimation algorithm to estimate
the source parameters, taking into consideration the large
amount of uncertainty in the measurements from the sensor
and the atmospheric dispersion process.
It is hypothesized that a more accurate and rapid response
could be achieved by planning the path of the UAS online in
response to the information gained during the mission. This is
termed an information-theoretic approach, which was shown in
[3] to outperform a systematic approach for STE in simulation.
The ability to estimate the source term of a dispersive release
using an information-theoretic approach was recently explored
in [4], where simple experiments validated the approach with a
ground robot equipped with a metal oxide sensor in an indoor
arena using fans to create a wind field and smoke to simulate a
HAZMAT release. Experiments in more realistic natural environments are limited. The one most related to this article is previous work we conducted [2]. A UAV executed a predefined
flight path to collect measurements from the onboard atmospheric sensor. The measurements were used to estimate the
source term of an atmospheric release using a sequential Monte
Carlo algorithm. The paper introduced a likelihood function to
relate the noisy measurements from the UAV, exacerbated by
the short sampling/averaging times used to take measurements,
with a simple atmospheric transport and dispersion (ATD)
model; this enabled the source term of a diffusive release to be
estimated in realistic outdoor conditions.
This article presents an extendable framework for information-theoretic hazardous source search and reconstruction followed by an experimental assessment of the system and then
realistic demonstrations conjured by potential end users from
the emergency services. The novelty of the work is in the integration of a recently proposed Bayesian STE algorithm with an

information-theoretic planner and a simplistic outline of a
framework for information-theoretic source reconstruction,
where each component can be extended to improve the overall
system. In addition, the
experimental assessment
of the information-theoHAZMAT sensor readings
retic search algorithm
using a UAV is the first of
will indicate the presence
its kind, paving the way to
a deployable system given
of HAZMAT, and this must
some further improvements to a few of the combe turned rapidly into a
ponents. The results show
a significant reduction in
warning to ensure the
search time by using the
proposed informationsafety of personnel
theoretic planner when
compared to that of a prein the vicinity.
planned uniform flight
path [2]. Finally, the value
of the system is demonstrated in realistic scenarios along with further demonstrations
of the inference engine using a manual and a preplanned flight
for data collection.
Related Work
There are many potential responses to releases of airborne
HAZMAT. Two prevailing methods are mapping of the hazard distribution [5]-[7] and localization of the source [8].
The focus of this article, however, is on source reconstruction,
otherwise known as STE, and the discussion on related work
is therefore limited accordingly.
The source term of a release includes its location and emission rate, in addition to other variables needed to model the
spread of the HAZMAT using an ATD model. The task is
most commonly formulated as an inverse problem and
approached using a network of static atmospheric concentration and meteorological sensors. Bayesian inference and optimization algorithms have both been proposed to estimate the
source term given the measurement data [1], along with alternative methods, e.g., neural network [9] and approximate
Bayesian computation [10]. Other sources of measurement
data have been seen in the literature, such as the use of toxic
health effects information [11].
Another option is to mount the atmospheric concentration sensors on mobile platforms, which in turn necessitates a
sensor planning function. Such a planning function has been
demonstrated in [12] to estimate the source terms of radiation releases using multiple robots. In other areas, methods
known as information-theoretic planning, information-based
planning, active information acquisition, active sensing, or
online informative path planning have been successfully
applied to such problems as simultaneous localization and
mapping (SLAM), known as active SLAM [13]; target search,
localization, and tracking [14], [15]; source localization [16];
and 3D mapping [17].
SEPTEMBER 2020

IEEE ROBOTICS & AUTOMATION MAGAZINE

109

IEEE Robotics & Automation Magazine - September 2020

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2020