2017 IEEE International Conference on Robotics and Automation (ICRA)

Singapore, May 29 - June 3, 2017

UB-ANC Planner: Energy Efficient Coverage Path Planning

with Multiple Drones
Jalil Modares, Farshad Ghanei, Nicholas Mastronarde and Karthik Dantu

Abstract— Advancements in the design of drones have led on the specific types of vehicles considered (e.g., ground
to their use in varied environments and applications such robots, indoor drones, or outdoor drones), their kinematics,
as battle field surveillance. In such scenarios, swarms of and the specific applications. Prior techniques have been
drones can coordinate to survey a given area. We consider
the problem of covering an arbitrary area containing obstacles optimized for shortest time to completion or control effi-
using multiple drones, i.e., the so-called coverage path planning ciency. However, a major challenge in the realization of such
(CPP) problem. The goal of the CPP problem is to find paths solutions is the limited energy on each drone.
for each drone such that the entire area is covered. However, We consider the problem of covering an arbitrary area con-
a major limitation in such deployments is drone flight time. taining obstacles using multiple UAVs/drones with a max-
To most efficiently use a swarm, we propose to minimize the
maximum energy consumption among all drones’ flight paths. min fair energy allocation across drones. Through experi-
We perform measurements to understand energy consumption mental measurements, we have determined that there are two
of a drone. Using these measurements, we formulate an Energy main factors that affect energy consumption in drones: dis-
Efficient Coverage Path Planning (EECPP) problem. We solve tance traveled and turns. Traditional coverage path planning
this problem in two steps: a load-balanced allocation of the algorithms, such as those based on the Traveling Salesman
given area to individual drones, and a minimum energy path
planning (MEPP) problem for each drone. We conjecture that Problem (TSP), are not ideal for drones because they only
MEPP is NP-hard as it is similar to the Traveling Salesman consider the distance traveled. We present a novel Energy
Problem (TSP). We propose an adaptation of the well-known Efficient Coverage Path Planning (EECPP) formulation that
Lin-Kernighan heuristic for the TSP to efficiently solve the explicitly considers the energy consumption characteristics of
problem. We compare our solution to the recently proposed drones in the path planning optimization, i.e., we not only
depth-limited search with back tracking algorithm, the optimal
solution, and rastering as a baseline. Results show that our consider the energy consumed traveling between consecutive
algorithm is more computationally efficient and provides more waypoints (similar to the TSP), but we also consider the
energy-efficient solutions compared to the other heuristics. energy consumed by the drone when it accelerates into and
out of turns. This work makes four contributions:
I. I NTRODUCTION
• From experimental measurements, we develop a linear
Networked unmanned aerial vehicles (UAVs) have model for energy consumption during drone flight.
emerged as an important technology for public safety, • Using this model, we formulate the EECPP problem and
commercial, and military applications including search and show that it is NP-hard.
rescue, disaster relief, precision agriculture, environmental • We decompose the EECPP problem into two sub-
monitoring, and surveillance. Many of these applications problems: a load-balancing problem that fairly divides
require sophisticated mission planning algorithms to coor- the area among drones and a minimum energy path
dinate multiple drones to cover an area efficiently. Such planning (MEPP) problem for each drone.
scenarios are complicated by the existence of obstacles, • We adapt heuristics proposed for solving the TSP to
such as buildings, requiring detailed planning for effective efficiently solve the MEPP sub-problem on each drone.
operation. Although a lot of work has been done on mission
The remainder of the paper is organized as follows. In
planning, optimal mission planning solutions depend heavily
Section II, we discuss related work. In Section III, we
describe our experimental energy measurements and the
ACKNOWLEDGEMENT OF SUPPORT AND DISCLAIMER: (a) energy model we derive from them. In Section IV, we
The University at Buffalo acknowledges the U.S. Government’s support in
the publication of this paper. This material is based upon work funded by introduce the EECPP problem formulation. In Section V, we
the US Air Force Research Laboratory under Grant No. FA8750-14-1-0073. present our simulation results comparing EECPP to rastering
(b) Any opinions, findings and conclusions or recommendations expressed (baseline), depth-limited search, and (where possible) an
in this material are those of the author(s) and do not necessarily reflect the
views of AFRL. optimal solution. We conclude in Section VI summarizing
F. Ghanei and K. Dantu were supported in part by the National Science our main results and future work.
Foundation under Grant Number IIS-1514395. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the II. R ELATED W ORK
author(s) and do not necessarily reflect the views of the National Science
Foundation. Recently, there has been a lot of work on coverage path
J. Modares and N. Mastronarde are with the Dept. of Elec- planning for UAVs. Ahmadzadeh et al. [1] introduce a
trical Engineering and F. Ghanei and K. Dantu are with the Dept. coverage algorithm for surveillance using a set of fixed-wing
of Computer Science and Engineering at the University at Buf-
falo, Buffalo NY, 14260, USA.{jmod, farshadg, nmastron, UAVs. They utilize dynamic programming to maximize the
kdantu}@buffalo.edu coverage of the area by a camera mounted on the drone.

978-1-5090-4633-1/17/$31.00 ©2017 IEEE 6182

Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
Maza et al. [7] propose a full coverage algorithm using a
set of heterogeneous UAVs (mostly helicopters). First, they
 

  
generate a polygonal partition of the area that takes into
account the capabilities of each individual UAV, such as flight
range. Each polygon in the partition is assigned to a UAV
that will cover it in a zig-zag pattern (i.e., a raster scan)
using a sweep direction that minimizes the number of turns.
Environmental obstacles are not considered in [1], [7].
Barrientos et al. [3] present an approach to cover an area
using multiple UAVs based on a depth-limited search with
back tracking. First, they present a task scheduler to partition
the target area into k non-overlapping areas for the k UAVs. Fig. 1: A UB-ANC Drone [8], [11] with a custom frame,
The partitioning procedure is based on a negotiation process waypoint-based Pixhawk flight controller, Raspberry Pi 2,
in which each UAV claims as much area as possible to cover. custom power sensor module, and 10 Ah battery. The drone
Then, the wavefront algorithm is used to cover each subarea. weighs 3kgs and achieves over 30 minutes of flight time.
Given that their application objective is similar to ours, we
have implemented a version of their algorithm and performed
comparisons with respect to computational time as well as custom frame, a waypoint-based Pixhawk flight controller, a
solution accuracy in the results. Raspberry Pi 2 for onboard computing, and a 6-cell 10 Ah
Di Franco et al. [5] discuss an energy-aware coverage battery that gives it over 30 minutes of flight time. Fig. 1
path planning solution for a single multi-rotor. They derive also shows the energy sensing module. This module consists
energy models for different operating conditions based on of four current sensing modules with ACS712 IC, which
real measurements. However, they only consider distance and translate the passing currents as analog output voltages. We
do not consider the impact of turns in their formulation. connected the power supply of the 4 motors to the sensors,
Torres et al. [12] propose a coverage path planning solu- and mounted all sensors together with an ADC converter
tion for 3D terrain reconstruction with a single UAV. They with ADS1115 IC and a logic level converter. The ADC has
decompose the area into one or more polygons and have four channels, each connected to one sensor, and can send
the UAV cover each polygon using a raster scan. They try the converted read values via I2 C. The logic level converter
to minimize the number of turns by calculating the optimal lets the IC communicate with the Raspberry Pi on the drone,
line sweep direction. which has a different logic voltage level.
Our work is focused on coverage path planning with mul- We used this instrumentation to measure the power con-
tiple UAVs. While most previous work attempts to generate sumption of the drone during flight to better understand the
paths with minimum distance, our proposed solution instead relationship between the distance, speed and direction change
optimizes energy consumption. Some work has explicitly of the drone and its total power consumption. Each exper-
attempted to minimize turns [3], [7], [12] . However, our iment was repeated multiple times and we show averaged
empirical results suggest that explicitly optimizing for energy results to alleviate anomalies from individual trials.
provides more energy-efficient paths than optimizing for • Straight Line Distance: In this test, we let the drone
turns or distance. fly in a straight line with a constant speed of 5m/s for
3 different distances: 20m, 40m, and 60m. We want
III. E NERGY C ONSUMPTION OF D RONE F LIGHT to see how the distance traveled affects the power
Consider a realistic scenario where a team of drones are consumption. During flight, the drone takes time to
commanded to survey an area. Such an area could contain ramp up to the desired velocity and starts slowing down
buildings, towers, and other man-made obstacles, or trees, prior to reaching the destination so it can come to a stop
hills, and other natural ones. Such obstacles can be convex at the destination waypoint.
or non-convex, making path planning fairly complex. In addi- • Effect of Velocity: In this test, we let the drone fly
tion, path planning for multiple drones to concurrently cover in a straight line for a constant distance of 40m, but
this area is even more challenging. Further, path planning of with 2 different target speeds: 5m/s and 10m/s. This
a single drone could be optimized for several objectives such is to understand the effect of the target flight speed on
as shortest time, least distance traveled, least energy used and the power consumption. As mentioned earlier, the drone
others. From empirical flight trials, we have concluded that must ramp up to its target speed and slow down prior
battery energy is the primary resource that limits flight times to reaching its destination.
of such aerial vehicles. Correspondingly, it would be ideal to • Effect of Turning: In this test, we want to observe
optimize paths based on energy consumption to enable the how direction changes affect power consumption. For
drones to cover the maximum area possible. this, we flew the drone 40m with a constant speed of
In order to better understand the power consumption 5m/s for five turn angles: 0◦ , 45◦ , 90◦ , 135◦ , and 180◦ .
dynamics of the drones, we equipped one of the drones with Specifically, for the 0◦ turn, we flew a 40m straight line
a power measurement module as shown in Fig. 1. It has a path and for the other turn angles we flew a 20m straight

6183
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
 




 
!

 

 
"

!




   
  


!



 


"


 
  



  

  

  
         

 
  


(a) Energy vs. Distance (b) Energy vs. Speed (c) Energy vs. Turn Angle
Fig. 2: Energy consumed as measured on the UB-ANC Drone for various patterns of flight. In Fig 2c, we omit the 180o
data point for line fitting as our planning does not allow re-visiting a node. It is shown here to demonstrate model validity.

line path, turned, and then flew another 20m straight path primitives (both distances and angles) and using them
line path. To isolate the energy consumption associated as parameters of the optimization. We now develop the
with turning, we subtracted the average straight path EECPP formulation based on our empirical power/energy
energy consumption from the average total energy that measurements on drone flight.
we measured for each angle.
Figs. 2a, 2b, and 2c show the average total energy con- IV. E NERGY E FFICIENT C OVERAGE PATH P LANNING
sumption with standard deviation for the distance, speed, F OR M ULTIPLE D RONES
and turn tests, respectively. Fig. 2a shows that increasing
Following from our energy measurements, we formulate
the distance traveled increases the energy consumed ap-
the Energy Efficient Coverage Path Planning (EECPP) prob-
proximately linearly. Assuming that traveling 0m incurs 0
lem in this section. The problem is divided into two sub-
energy cost, we performed a linear fit on the measured
problems: (i) fairly dividing the given area among the drones,
data and determined that the drone consumes approximately
and (ii) minimum energy path planning (MEPP) for each
λ = 0.1164kJ/m. Fig. 2b shows the relationship between
drone. Our formulation allows drones to start in different
the energy consumption and the target speed. The drone
locations and requires each drone to return to its starting
was flown for a distance of 40m and commanded to fly at
point akin to the TSP. These assumptions are drawn from
the given speed. Lower speeds result in greater time spent
intuition from real applications where surveying is part of a
by the drone in the air and correspondingly greater energy
larger operation.
consumption. Fig. 2c shows the effect of the turn angle on
the energy. It is interesting to observe that increasing the turn
A. Problem Modeling
angle increases the energy consumed for the same distance
traveled in an almost linear manner. It is also noteworthy As our measurements have shown in Section III, a drone’s
that the variance in energy consumed grows with greater energy consumption depends on the distance it travels and the
turn angle. We performed a linear fit on the measured data number and degree of turns in its path. To find the minimum
and determined that the drone consumes approximately γ = energy path for each drone, we formulate a vehicle routing
0.0173kJ/deg. This suggests that intelligently reducing the problem (VRP) [4], which is a more general case of the
number and magnitude of turns in a path can potentially multiple traveling salesman problem (mTSP). The original
reduce energy consumption. Note that we use our measured VRP problem is a min-sum optimization, which tries to
values of λ and γ when we solve the optimizations proposed minimize the sum of costs over all vehicles. We adapt that
in Section IV-B. to a min-max formulation where we want to minimize the
Figs. 3a, 3b, and 3c show the power consumption and maximum cost incurred (energy expended) by any drone. In
flight times for the same tests. All three graphs show that the literature, this has been referred to as the Newspaper Routing
average power consumption is nearly constant for all traveled Problem [2], which considers fairness among all vehicles
distances, speeds, and turn angles. Therefore, the difference (agents). However, our problem differs from the Newspaper
in energy consumption across tests is primarily due to the Routing Problem because the objective function not only
duration of the test. Based on the flight controller’s design, depends on the distance traveled, but also on the turns.
the drone often slows down considerably when it enters a Surveillance of a given area requires coverage of all
turn, which results in an increased flight time (Fig. 3c). locations. However, given that the drone is flying at a given
We note that for a different choice of drone, the energy height, it is able to view a large area from its vantage point.
consumption patterns might be different, but we believe Therefore, we represent the area to be surveyed as a set of
that the trends are reasonably indicative of the energy grid cells and assume that a grid cell is covered if the drone
consumption in most drones of this size. Our formulation visits its center. Formally, we represent the grid as a graph
allows for modeling the energy consumed for each of the G (V , E ), where V is the set of nodes and E is the set of

6184
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
!



 

 !



 

 !



 



 !   !   ! 


!


!



!

  











     
  
  
  
  
  
  
     
       

 
 


(a) Power and Time vs. Distance traveled (b) Power and Time vs. Speed of travel (c) Power and Time vs. Turn Angle
Fig. 3: Average power draw and time for different missions.

edges. We let i, j, k ∈ V denote a specific node and ei j ∈ E

denote an edge between nodes i and j. 
We define the pair (αi , βi ) as the Cartesian coordinate of
node i ∈ V and let ci j denote the cost of traversing the edge
ei j between node i ∈ V and node j ∈ V . In our path planning
 

optimization, we assume that drones traverse adjacent cells.
In other words, ei j ∈ E if nodes i, j ∈ V are adjacent and ei j ∈
/
E otherwise. Based on our measurements in Section III, we
assume that the cost (energy) is proportional to the distance
traveled: i.e.,
 
λ (αi − α j )2 + (βi − β j )2 , if ei j ∈ E Fig. 4: A cell and its neighbors, and the exterior angle for
ci j = (1)
∞, otherwise. three nodes on the path.
In other words, if two nodes are adjacent, then the energy B. Problem Formulation
cost to traverse the edge between them is proportional to the
Euclidean distance between the centers of their correspond- Let A denote the set of agents that will cover the area
ing cells (where the parameter λ kJ/m is specific to the drone and let va ∈ V denote the starting node for agent a ∈ A . We
as described in Section III); however, if two nodes are not indicate which edges each agent traverses using the binary
adjacent, then the cost to traverse them is infinite (i.e., it is decision variable xiaj ∈ {0, 1}, where
not possible to directly traverse the two nodes because they 
are not connected by any edge). 1, if agent a ∈ A traverses edge ei j ∈ E
Let θi jk denote the exterior angle between nodes i, j, k ∈ V xiaj = (7)
0, otherwise.
(Fig. 4). The squared length of the edges of the triangle made
by nodes i, j, k can be determined as follows: Given a feasible path assignment (i.e., a sequence of edges),
r = (αi − α j )2 + (βi − β j )2 , (2) agent a ∈ A will incur a total cost of
s = (α j − αk )2 + (β j − βk )2 , and (3)
∑ ∑ ci j xiaj + ∑ ∑ ∑ qi jk xiaj xajk , (8)
2
t = (αk − αi ) + (βk − βi ) 2
(4) i∈V j∈V \{i} i∈V j∈V \{i,va } k∈V \{ j}

We know that given the lengths of three sides of a triangle, where the first term in the cost function is proportional to
we can calculate an internal angle using the Law of Cosines. the distance traveled (as in the TSP) and the second term is
It follows that the exterior angle between nodes i, j, k ∈ V proportional to the sum of turn angles (unique to the EECPP).
can be written as: Formally, the EECPP problem can be stated in Equa-
 
−1 (r + s − t) tions 9. The objective of the EECPP (Equation (9a)) is
θi jk = π − cos √ radians. (5)
4rs to determine the paths for each drone that minimize the
From our empirical energy measurements, we model the maximum cost incurred by any individual drone, where the
cost associated with a feasible turn, denoted by qi jk , to be cost function is defined in Equation (8). There are several
proportional to the angle of the turn (where the parameter γ constraints governing this optimization. First, each node
kJ/deg is specific to the drone as described in Section III): should be visited exactly once (Equation (9b)). Next, we
 need flow conservation constraints which ensure that, once
γ 180
π θi jk , if ei j , e jk ∈ E a drone visits a node, it also departs from the same node
qi jk = (6)
∞, otherwise. (Equation (9c)). Third, we incorporate extensions of MTZ-
based SECs [4] (subtour elimination constraints) to a three-

6185
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
index model (Equation (9d)).1 Finally, ui is a dummy variable D. Sub-Problem 2: Minimum Energy Path Planning (MEPP)
associated with node i ∈ V which is assigned by the solver.
The second sub-problem is the minimum energy path
planning (MEPP) problem. After dividing the area in sub-
min max
a∈A
∑ ∑ ci j xiaj + ∑ ∑ ∑ qi jk xiaj xajk problem 1, we use mixed integer quadratic programming
i∈V j∈V \{i} i∈V j∈V \{i,va } k∈V \{ j} (MIQP) to formulate the MEPP problem. The problem is
(9a) very similar to the EECPP (9a) except that there are no
s.t. ∑ ∑ xiaj = 1, ∀ j ∈ V (9b) indices for individual drones. It is shown below.
a∈A i∈V \{ j}

∑ xiaj − ∑ xajk = 0, ∀a ∈ A , ∀ j ∈ V min ∑ ∑ ci j xi j + ∑ ∑ ∑ qi jk xi j x jk

i∈V \{ j} k∈V \{ j} i∈V j∈V \{i} i∈V j∈V \{i,va } k∈V \{ j}
(9c) s.t. ∑ xi j = 1, ∀ j ∈ V
ui − u j + |V | xiaj ≤ |V | − 1, i∈V \{ j}

∀a ∈ A , ∀i, j ∈ V \{va } and i = j (9d) ∑ xi j = 1, ∀i ∈ V

j∈V \{i}
ui ∈ Z , ∀i ∈ V (9e)
ui − u j + |V | xi j ≤ |V | − 1,
xiaj ∈ {0, 1}, ∀i, j ∈ V and ∀a ∈ A (9f)
∀i, j ∈ V \{va } and i = j
ui ∈ Z , ∀i ∈ V
The problem shown above is an NP-hard mixed integer
xi j ∈ {0, 1}, ∀i, j ∈ V
quadratic constrained program (MIQCP); therefore, it does
not scale well beyond a few drones and dozens of cells. To
We conjecture that the above minimum energy path planning
overcome this limitation, we decompose the problem into
problem is NP-hard since it is similar to the TSP, but with
two sub-problems: the first sub-problem assigns a set of
additional quadratic terms to account for the turning costs.
cells to each drone and the second determines the minimum
To solve this problem efficiently when a large number of
energy path that each drone will follow to cover these cells.
nodes are assigned to a drone, we propose a modification of
C. Sub-Problem 1: Load Balancing (LB) the well-known Lin-Kernighan Heuristic (LKH [6]). While
The first sub-problem is the so-called load balancing (LB) the conventional LKH only considers the distance traveled,
problem. We use mixed integer linear programming (MILP) we modify it to also account for the drone’s turning costs.
to divide the nodes among the users with linear complexity. Lin-Kernighan Heuristic for Drones (LKH-D): We now
Let cai denote the distance between agent a ∈ A and node briefly describe the LKH used to solve the TSP. We then
i ∈ V . Let xai denote a decision variable that is set to 1 show how we have modified the LKH to account for the cost
if node i ∈ V is assigned to agent a ∈ A and is set to 0 of turns in the MEPP problem. We call the new algorithm
otherwise: i.e., LKH for Drones (LKH-D).
 LKH begins by determining a feasible tour T that visits
1, if node i ∈ V is assigned to agent a ∈ A each node exactly once and returns to the origin node. The
xai = (10)
0, otherwise. tour T is associated with a cost f (T ), which is equal to the
length of the tour. LKH works by iteratively improving the
Based on the starting positions of the drones, we would
initial tour using a specific transformation (see, e.g., [6]).
like to assign grid cells to them to minimize the maximum
After applying the transformation on the tour T , a new
energy incurred across them. This can be formulated as:
feasible tour T  is obtained, which has a cost f (T  ). If the
min max
a∈A
∑ cai xai (11a) gain g(T, T  ) = f (T ) − f (T  ) is positive (i.e., the tour T  has
i∈V a lower cost than the tour T ), then the new tour is adopted;
s.t. ∑ xai = 1, ∀i ∈ V (11b) otherwise, it is thrown away. This procedure is repeated
a∈A iteratively until a specific stopping condition is met (see,
xai ∈ {0, 1}, ∀a ∈ A and ∀i ∈ V (11c) e.g., [6]). Fig. 5a and Fig. 5b show a four node tour T and
a feasible tour T  obtained by an appropriate transformation
This can be solved by a linear program with complexity that
on T : in this example, T  is obtained from T by replacing
is linear in the product of the number of drones and the
edges BD and CA with edges BC and DA, respectively.
number of grid cells.
Our proposed LKH-D algorithm follows the same iterative
approach as the conventional LKH, but uses a different cost
function to account for the drone’s turning costs. Specifically,
f (T ) is calculated as a weighted sum of the length of the
tour and the sum of the turn angles within the tour (these
1 A subtour is a closed path that starts from one node and returns to that
are the four exterior angles illustrated in Figs. 5a and 5b.) In
node. The subtour elimination constraints prevent the optimization solver other words, f (T ) is equal to the energy cost defined in (8)
from returning undesirable subtours as solutions. associated with the tour T .

6186
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
  $##

"

 










(b) New feasible tour T  obtained 

(a) Initial feasible tour T . from tour T .



% % % %
Fig. 5: Illustration of the Lin-Kernighan Heuristic (LKH).
!
&"

V. S IMULATION R ESULTS (a) Computation time

"%*
#"&(!$'#"

We perform several sets of simulations to demonstrate the


benefits of our proposed heuristics for the EECPP problem
with multiple drones and the MEPP problem (i.e., sub- 

problem 2) for a single drone. We compare our solution 

with (a) simple rastering, which has no planning cost, (b) 
&'%
a previously proposed depth-limited search (DLS) algorithm
with backtracking for multi-robot search [3],2 and (c) an 

brute-force optimal solution wherever possible. (We imposed 

a 1 hour execution time limit for all algorithms, and present 

the best result obtained in that time). We have compared

these algorithms in several scenarios to demonstrate the ) ) ) )
utility of our solution. %
+ &
For all our simulations, we use the UB-ANC Emulator [9], (b) Performance
[10],3 which is an emulation framework that we developed
to simplify the transition from simulation to experimentation Fig. 6: Run-time (max: 1hr) and performance of MEPP on
with multi-agent drone systems. In particular, our simulations rectangular grids of different sizes without obstacles.
leverage a software-in-the-loop (SITL) version of the flight
controller’s firmware, which allows our simulations to be rectangular maps with dimensions 2 × 4, 3 × 6, 4 × 8, and
immediately ported to actual hardware. 5 × 10. The cells in each of these maps are 10m × 10m
We perform comparisons with the benchmark algorithms squares and there are no obstacles. Fig. 6a shows how the
on several dimensions. We vary the area of coverage by a computation time scales with the number of cells for each
single drone to compare the scalability of the algorithms algorithm and Fig. 6b shows the energy consumption under
with increased area (number of grid cells). We compare each algorithm. Please note that the time axis in Fig. 6a is in
the efficiency of each approach by comparing the total log-scale. As would be expected, the optimal solution for the
energy consumed by the drone for that mission. We also MEPP problem, which is similar to the TSP, is computation-
perform these comparisons in four scenarios with one or ally expensive. Rastering does not require any planning and
more obstacles as shown in Fig. 7. Finally, we demonstrate is not represented in the figure. In comparison to DLS [3],
a full run of our multi-robot path planning problem by the proposed LKH-D is three orders of magnitude faster for
simulating a set of drones covering a large area. Our results the large map. Fig. 6b shows that for small areas without
are from simulations on a standard laptop. We impose a 1-hr obstacles all algorithms achieve comparable performance to
run limit on all our algorithms. the exhaustive CPLEX solution (which is optimal for grid
We note that Sections V-A and V-B focus primarily on sizes up to 4 × 8).
the MEPP problem for a single drone. In Section V-C,
we consider the full EECPP problem with multiple drones B. Energy Efficiency and Algorithm Adaptivity
covering a large area. As discussed in the introduction, a major challenge in
A. Algorithm Scalability path planning is adapting to real-world constraints such as
obstacles and areas shaped in a non-standard manner. To
First, we show how the compared algorithms perform in understand the effects of obstacles on the computation time
terms of computation time and energy efficiency for simple and performance of the compared algorithms, we generated
four 8 × 15 rectangular maps as illustrated in Fig. 7. Given
2 Since no source code was available for this algorithm, we have
the size of the area (120 cells), we were unable to run the
faithfully implemented our version of the proposed algorithm and verified
its functionality with results from various papers on this algorithm. CPLEX to solve the MEPP problem to completion in all
3 https://github.com/jmodares/UB-ANC-Emulator cases. Instead, we run the optimization for one hour and

6187
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
 (a) Area 1 (b) Area 2 (c) Area 3 (d) Area 4 (a) Rastering (b) DLS (c) CPLEX (d) LKH-D
Fig. 7: Rectangular grid maps used to evaluate the algo- Fig. 9: Visualization of the planned paths for different
rithms. Grey cells represents obstacles. algorithms in area 4 of Fig. 7. The drone’s tour starts and
ends in the upper-right corner of the grid.




  TABLE I: Simulation statistics for area 4 in Fig. 7.

  Algorithm Raster DLS (1 hr) CPLEX LKH-D

  Energy (kJ) 162.2 156.2 127.7 127.7
Comp Time (s) 0 3600 1276 0.4

 Distance (m) 1244.8 1043.8 994.1 994.1
Turns (degree) 1620 2970 1620 1620



 Fig. 9 illustrates the planned paths under each algorithm
for the area shown in Fig. 7d, and Table I provides some


statistics about these paths (energy, computation time, total

 
 
 


 distance, and total degree of turns). The proposed heuristic
(LKH-D) achieves the optimal energy consumption, the
(a) Computation time minimum distance traveled, and the minimum total turns,
and is executed in less than half a second.
!%
"$ #





C. Multi-Drone Path Planning

To show the application of the proposed method for


solving the EECPP problem (load balancing followed by

  LKH-D to solve the MEPP) in a large-scale scenario, we
  executed it with a varying number of drones (5, 10, 15,


and 20) on the University at Buffalo’s North Campus in


simulation. A top view of the area can be seen in Fig. 11.

"#! The area is decomposed into over 3000 cells measuring

20m × 20m. Running the multi-agent coverage path planning
!
 !
 !
 !

"#
" on the area with different numbers of drones, with both
LKH and LKH-D algorithms, we obtain the results shown
(b) Performance in Fig. 10a and 10b. Fig. 10a compares them with respect
Fig. 8: Comparison of algorithms over areas in Fig. 7. to average computation time required for the path planning
per drone across all drones in each mission. Fig. 10b
compares these algorithms with respect to average energy
expended per drone. As expected, increasing the number of
report results returned by the CPLEX optimization solver. available drones allows each agent to cover less area, and
Fig. 8a shows the computation time for the compared correspondingly decreases the computation time and energy
algorithms when they are applied to the four areas defined expended for the mission. Finally, LKH-D achieves lower
in Fig. 7, and Fig. 8b shows the corresponding energy energy consumption than LKH, at the expense of increased
consumption. In all scenarios, our proposed heuristic (LKH- computation time, because it considers the turning costs.
D) performs at least as well as the time-limited CPLEX
solution and does approximately 15-20% better than the VI. C ONCLUSION
DLS and rastering solutions. Moreover, LKH-D achieves this We formulated the Energy Efficient Coverage Path Plan-
performance in 3-orders of magnitude less time than we ning (EECPP) problem for covering an arbitrary area con-
allowed for the CPLEX and DLS solutions. These results taining obstacles using multiple drones. The goal of the
highlight the benefits of both our modeling and the proposed EECPP problem is to minimize the maximum energy re-
heuristic for energy-efficient path planning. quired for any individual drone to traverse its assigned

6188
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.
 !
 

 

 





  
 

 Fig. 11: UB North Campus. Areas dense with buildings are
assumed to obstacles (shaded with diagonal lines).
(a) Computation
time
!
"
 


previously proposed algorithms and comes up with more

 energy-efficient paths.



 In future work, we plan to extend the proposed frame-
work to integrate network performance metrics into the


path planning optimization, e.g., to optimize connectivity,
throughput, or delay. We implemented UB-ANC Planner



based on the algorithms mentioned in the paper. UB-ANC
Planner is available as open-source at https://github.


com/jmodares/UB-ANC-Planner.

R EFERENCES
   [1] Ahmadzadeh, A., Keller, J., Pappas, G., Jadbabaie, A., Kumar, V.: An
 

 optimization-based approach to time-critical cooperative surveillance
and coverage with uavs. In: Experimental Robotics, pp. 491–500.
(b) Performance Springer (2008)
[2] Applegate, D., Cook, W., Dash, S., Rohe, A.: Solution of a min-max
Fig. 10: Average path planning computation time (per drone), vehicle routing problem. INFORMS Journal on Computing 14(2),
and energy consumption (per drone) for a set of drones 132–143 (2002)
[3] Barrientos, A., Colorado, J., Cerro, J.d., Martinez, A., Rossi, C.,
covering UB North campus. Comparison between LKH and Sanz, D., Valente, J.: Aerial remote sensing in agriculture: A practical
LKH-D algorithms. approach to area coverage and path planning for fleets of mini aerial
robots. Journal of Field Robotics 28(5), 667–689 (2011)
[4] Bektas, T.: The multiple traveling salesman problem: an overview of
formulations and solution procedures. Omega 34(3), 209–219 (2006)
path. Unlike the conventional multiple traveling salesman [5] Di Franco, C., Buttazzo, G.: Energy-aware coverage path planning of
problem (mTSP), the vehicle routing problem (VRP), and uavs. In: Autonomous Robot Systems and Competitions (ICARSC),
the newspaper routing problem, which only consider the 2015 IEEE International Conference on, pp. 111–117. IEEE (2015)
[6] Lin, S., Kernighan, B.W.: An effective heuristic algorithm for the
distance traveled by each agent, we accounted for the energy traveling-salesman problem. Operations research 21(2), 498–516
consumption characteristics of drones in our optimization. In (1973)
particular, we included a term to account for the additional [7] Maza, I., Ollero, A.: Multiple uav cooperative searching operation
using polygon area decomposition and efficient coverage algorithms.
energy consumed by drones when they accelerate into and In: Distributed Autonomous Robotic Systems 6, pp. 221–230. Springer
out of turns. This decision was justified by experimental (2007)
energy measurements on an actual drone. However, we con- [8] Modares, J., Mastronarde, N.: Ub-anc: A flexible airborne networking
and communications testbed: Poster. In: Proceedings of the Tenth
jectured that this problem is NP-hard, and we decomposed ACM International Workshop on Wireless Network Testbeds, Exper-
it into two sub-problems. The first sub-problem divides the imental Evaluation, and Characterization, WiNTECH ’16, pp. 95–96.
area among drones and has linear complexity. The second ACM, New York, NY, USA (2016)
[9] Modares, J., Mastronarde, N., Dantu, K.: Ub-anc emulator: An
sub-problem then determines the path for each drone to emulation framework for multi-agent drone networks. In: IEEE
cover its assigned area. Although the second sub-problem is International Conference on Simulation, Modeling, and Programming
similar to the TSP (which is NP-hard), we believe that this is for Autonomous Robots (SIMPAR) (2016)
[10] Modares, J., Mastronarde, N., Dantu, K.: Ub-anc emulator: An
acceptable as long as there are enough drones to divide the emulation framework for multi-agent drone networks: Demo. In:
areas into reasonably sized regions (with under 50 cells) for Proceedings of the Tenth ACM International Workshop on Wireless
the solver to process in a timely manner. More complex grids Network Testbeds, Experimental Evaluation, and Characterization,
WiNTECH ’16, pp. 93–94. ACM, New York, NY, USA (2016)
can be solved using sub-optimal heuristic algorithms. To this [11] Modares, J., Mastronarde, N., Medley, M.J., Matyjas, J.D.: Ub-anc:
end, we adapted a heuristic for the TSP problem (LKH) An open platform testbed for software-defined airborne networking
and proposed the LKH-D algorithm incorporating energy and communications. arXiv preprint arXiv:1509.08346 (2015)
[12] Torres, M., Pelta, D.A., Verdegay, J.L., Torres, J.C.: Coverage path
consumption into the solution. We showed in simulation planning with unmanned aerial vehicles for 3d terrain reconstruction.
that our proposed heuristic is computationally faster than Expert Systems with Applications 55, 441–451 (2016)

6189
Authorized licensed use limited to: University at Buffalo Libraries. Downloaded on March 17,2025 at 17:38:18 UTC from IEEE Xplore. Restrictions apply.