Robotics and Autonomous Systems 98 (2017) 158–173

Contents lists available at ScienceDirect

Robotics and Autonomous Systems

journal homepage: www.elsevier.com/locate/robot

Two-layer obstacle collision avoidance with machine learning for

more energy-efficient unmanned aircraft trajectories
Youngjun Choi *,1 , Hernando Jimenez 1 , Dimitri N. Mavris 2
Georgia Institute of Technology, Atlanta, GA, 30332-0150, United States

highlights

• Novel optimal obstacle avoidance algorithm with two-layer structure.

• Global-path optimization resolves multiple obstacles via a clustering algorithm and specifies a flight path.
• Local-path optimization is defined as a multi-phase optimal trajectory problem.
• The proposed two-phase collision avoidance algorithm is demonstrated and shown to yield more energy-efficient collision avoidance trajectories in a
simulated multi-obstacle environment when compared to a traditional one-layer obstacle algorithm.

article info a b s t r a c t
Article history: This paper proposes a new two-layer obstacle avoidance algorithm that allows an unmanned aircraft
Received 6 October 2015 system to avoid multiple obstacles with minimal effort. The algorithm includes a global-path optimization
Received in revised form 6 July 2017 that identifies the number of obstacles resulting from a clustering technique based on obstacle infor-
Accepted 3 September 2017
mation from an airborne sensor, and specifies a potential threat. A local-path trajectory optimization
Available online 18 September 2017
employs a model predictive control structure based on a multi-phase optimal trajectory resulting from
approximated dynamics, vehicle constraints, and the result of the global-path optimization. Numerical
Keywords:
Obstacle avoidance flight simulations are conducted with a conventional one-layer obstacle avoidance algorithm and the
Optimal trajectory two-layer obstacle avoidance algorithm. The results of the numerical simulation show that the proposed
Clustering algorithm two-layer optimal obstacle avoidance algorithm generates more energy-efficient avoidance trajectories
Model predictive control when an unmanned aircraft meets multiple obstacles.
UAV © 2017 Elsevier B.V. All rights reserved.
Path-planning

1. Introduction The current separation assurance paradigm for unmanned air-

Unmanned aircraft and their integration into the airspace sys-
tem have gained significant attention and popularity given the eco-
nomic potential stemming from diverse applications, such as par-
cel delivery, support of rescue missions, and surveillance missions
using various platforms. However, the prospect of fully integrated
airspace operations is limited by technological and regulatory chal-
lenges, including separation assurance, human system integration
and communication. Among core challenges, detect and avoid has
been recognized as a key enabler for separation assurance by key
stakeholders such as the Federal Aviation Administration (FAA) [1].
* Corresponding author.
E-mail addresses: ychoi95@gatech.edu (Y. Choi),
hernando.jimenez@asdl.gatech.edu (H. Jimenez),
dimitri.mavris@aerospace.gatech.edu (D.N. Mavris).
1 Research Engineer, School of Aerospace Engineering, Georgia Institute of
Technology.
2 S.P. Langley Distinguished Regents Professor, School of Aerospace Engineering,
Director Aerospace Systems Design Laboratory, Georgia Institute of Technology.
craft builds upon that for manned traffic as a multi-layered frame-
work with overlapping capabilities that include operating pro-
cedures, strategic separation, tactical separation, self separation,
and collision avoidance, following progressively shorter time hori-
zons [2,3]. Strategic separation, for instance, is exercised by air traf-
fic management services and addresses conflict resolution for time
horizons roughly between 3 min and 10 min. Collision avoidance
is realized by on-board systems for conflict resolution within ap-
proximately 15 s. The detect and avoid capability of an unmanned
aircraft is a key enabler for self-separation and collision avoidance,
and it contributes to top-level operational capabilities in fully
integrated operations given its coupling with other critical ele-
ments, such as sensing/surveillance, communications, functional
allocation, and human factors.
There are numerous studies investigating collision avoidance
concepts and techniques, which can be generally divided into
five groups: stochastic methods, road map methods, potential
field approaches, geometric methods, and optimization-based

http://dx.doi.org/10.1016/j.robot.2017.09.004
0921-8890/© 2017 Elsevier B.V. All rights reserved.
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
methods. Stochastic methods effectively search nonconvex high-
dimensional spaces for global or local obstacle avoidance paths
based on the environment as perceived by airborne sensors. The
rapid random tree (RRT) [4,5] is a popular method of this kind. Road
map methods use visual graph characteristics and path-planning
algorithms to partition collision-free paths and create a piecewise
linear path or curved path using smoothing techniques [6]. The
Theta∗ algorithm has recently been introduced to overcome the
restricted heading angle issue of the classical A∗ algorithm to solve
the 3D path-planning problem in an urban environment [7].
The fundamental concept of artificial potential field methods
is the creation of a force map where a waypoint generates an
attractive force and obstacles generate repulsive forces. Based on
the result of the force map, an algorithm generates an optimal
collision-free path. The artificial potential field has been intro-
duced by Khatib et al. to solve an obstacle avoidance problem for
robotics applications [8]. This approach has the benefits of a simple
structure and ease of implementation but also has some impor-
tant shortcomings, such as local minima where the total force is
close to zero and non-reachable waypoints due to proximity to
obstacles [9–11]. Improvements to address these issues have been
proposed, for instance in work by Li et al. [12] and Mac et al. [11],
by introducing additional force terms and condition exceptions.
Among geometric methods, one approach uses the point of
closest approach for a collision avoidance problem against a mov-
ing airplane. This technique specifies a conflict condition based
on the vehicles’ geometric relationship, and the two aircraft ma-
neuver away from the conflict region through a vector sharing
resolution [13]. Another geometric approach utilizes a collision
cone to yield a closed-form solution for an aiming point [14]. The
collision cone approach has some inherent shortcomings for which
solutions have been developed in more recent work [15].
Optimization-based and geometric methods are closely re-
lated, and have presented a range of promising solutions. For
instance, Chakravarthy and Ghose proposed a collision cone tech-
nique to avoid a moving obstacle with an irregular shape in
two-dimensional space [16]. Watanabe et al. extended this ap-
proach with a minimal-effort optimization framework in three-
dimensional space [17]. Another example is that of Schouwe-
naars et al. [18] who suggested mixed-integer linear programming
(MILP) that incorporates binary constraints based on the area infor-
mation outside of an obstacle. Yoshiaki et al. [19] propose a more
precise MILP algorithm for a three-dimensional collision avoid-
ance problem that includes the construction of a coarse cost map
and a detailed trajectory optimization. These MILP frameworks
have been employed for a real-time collision avoidance problem
in diverse platforms [20]. However, solving MILP is commonly
implemented via the branch and bound method, which is compu-
tationally expensive [20,21]. Another optimization-based method,
proposed by Moon et al. [22], solves for an optimal avoidance
trajectory based on a safe position around an obstacle determined
from on-board sensor information and a flight envelope protection
function. Concurrent detection of multiple proximate obstacles
results in them being mishandled as a single obstacle, even when
there is sufficient clearance to fly safely between them, resulting
in highly inefficient trajectories. Kang et al. [23] expanded this
framework to by incorporating external surveillance sources with
obstacle information. Although some improvements are observed,
this concept is vulnerable to surveillance data loss and essentially
reduces to the same local-optimization approach it seeks to im-
prove upon. In this paper we address the question of how an
optimal collision avoidance algorithm can produce highly energy-
efficient trajectories in a multi-obstacle environment while relying
solely on on-board sensors and capabilities.
This work proposes a two-layer obstacle collision avoidance
algorithm that incorporates a global-path optimization and a local-
path optimization. In the global-path optimization, based on the
159
detected obstacle information from an airborne sensor, an on-
board system identifies the number of obstacles through a cluster-
ing technique. This clustering routine solves a distance-based con-
straint optimization problem that includes minimum-separation
distance between pairs of adjacent obstacles. Then, based on the
cluster results and the velocity information, the on-board system
judges a potential threat among the identified multiple obstacles.
The clustering algorithm has been successfully applied to resolve
multiple obstacles as separate objects in the field of ground mobile
robots [24]. However, the UAV collision avoidance problem is a
higher dimensional clustering problem that requires more com-
putational resources to handle three-dimensional airborne LiDAR
information. Therefore, this paper also explores diverse techniques
and compares them to achieve a more computationally favorable
approach. In the local-path optimization, we employ a multi-phase
optimal obstacle avoidance problem.
Many clustering techniques have been introduced because of
their simplicity and scalability for a broad spectrum of applications.
The k-means algorithm, one of the most popular and classical
clustering methods, partitions a data set into k clusters by sequen-
tially assigning points to clusters based on their proximity and
updating the cluster definitions to reflect said point assignment.
However, the k-means algorithm requires the specification of the
number of clusters k a priori. To circumvent this issue, several
approaches have been proposed, for instance, by evaluating some
top-level partition quality metric for which said metric is optimal
(see for example Ref. [25]). These approaches often fail to provide
sufficiently tractable results and are computationally expensive.
Jia et al. posit that if the point cloud of each group has a non-
convex shape, these clustering algorithms yield results that only
guarantee a local minimum [26]. Another clustering algorithm is
the NJW algorithm proposed by Ng et al. as a classical spectral
clustering method, which is computationally tractable [27]. In this
approach, the number of clusters k can be determined from the
top k eigenvalues of the Laplacian matrix. However, in that case,
determining the number k of clusters does not have a tractable
standard, and the top k eigenvectors do not always produce correct
clustering results [26].
For the obstacle avoidance trajectory problem, the number of
clusters k cannot be user-specified. Moreover, resolution of clus-
ters as separate objects must be defined according to an explicitly-
defined minimum separation distance. To solve the issue of the
NJW algorithm, the number of clusters k can be determined from k-
block diagonality of the Laplacian matrix [28]. Another clustering
technique that does not need any a priori information of k is DB-
SCAN, proposed Ester et al. [29], a density-based spatial clustering
algorithm. In DBSCAN, the number of clusters is not defined a priori
but rather results from the clustering process itself, driven by the
grouping of points that meet a user-defined minimum density.
For the clustering algorithm in the global-path optimization, we
select two clustering algorithms as candidate solutions and com-
pare their performance with respect to clustering accuracy and
computational efficiency.
For the online trajectory optimization, we employ a model
predictive control (MPC) scheme that produces computationally
feasible solutions. That is, solving actual nonlinear constrained
dynamics trajectory optimization is an NP-hard problem, but MPC
solves approximated linear dynamics and simplified constraints,
and regularly updates the results of the optimal trajectory in a
real-time manner. This MPC approach provides a computationally
tractable solution for an online optimal collision avoidance prob-
lem and has been successfully implemented in the past [22,23,30].
We hypothesize that our two-layer collision avoidance algo-
rithm yields more energy efficient trajectories without incurring
a prohibitive computational burden relative to the single-layer
approach. In the remainder of this paper, we first present the
formulation and implementation of our algorithm, and then test
the hypothesis via direct comparison of numerical simulations.
160 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
2. New path planning architecture using a learning algorithm
2.1. Two-level algorithm concept
The multi-obstacle avoidance problem is inherently challeng-
ing. In general, the mathematical complexity and computational
expense of an optimal trajectory solution grows quickly with the
explicit treatment of multiple obstacle constraints. It follows that
most optimal trajectory frameworks in recent work do not explic-
itly consider multi-obstacle scenarios [22,23,31]. Instead, existing
algorithms employ the collision avoidance framework for one ob-
stacle regardless of the number of downstream trajectory threats,
producing inefficient trajectories whenever safe flight between
obstacles is feasible. To address this major shortcoming, the new
algorithm proposed here features a two-layer structure accommo-
dating online global-path and local-path optimization processes,
illustrated in Fig. 1.
The global-path optimization uses on-board sensor data to ef-
ficiently resolve multiple downstream obstacles based on their
relative location and separation. It identifies the potential threat
among the resolved obstacles. The local path optimizer produces
an optimal collision avoidance trajectory using the information of
the potential threat. To do so, it solves a multi-phase optimal tra-
jectory problem based on vehicle dynamics, constraints, obstacle
information, and mission waypoints. Two fundamental phases are
defined for this problem: obstacle avoidance and recovery. In the
obstacle avoidance phase, the terminal trajectory point is an in-
termediate waypoint that satisfies three linear constraints around
the obstacle, one above and one on either side, guaranteeing a
minimum safe separation distance between the aircraft and the
obstacle. In the recovery phase, the trajectory takes the aircraft
from the intermediate waypoint to a prescribed target, or final
waypoint. The avoidance path is continuously optimized based on
updated airborne sensor information and vehicle states.
2.2. Two-level algorithm in the guidance, navigation, and control
architecture
The guidance navigation and control architecture employed in
this work, illustrated in Fig. 2, is consistent with typical constructs
in current practice. Optimal trajectory generation realized with
guidance and navigation elements, as well as application of air-
craft controller to the flight dynamics, ensures adherence to said
trajectory. In our two-level concept, guidance and navigation have
an off-line block where the trajectory is defined a priori in accor-
dance with a prescribed mission profile, pre-defined waypoints,
and available obstacle information (if applicable or available). The
on-line counterpart regularly updates the trajectory based on the
current vehicle state and sensor data to compensate for perturba-
tions or the avoidance of unknown obstacles. The global and local
components of the proposed trajectory optimization concept are
realized in this region of the architecture. As shown in Fig. 2, sensor
data corresponding to detected obstacles in the operating environ-
ment are passed to the global block where distinct obstacles are
resolved, and potential threats are specified.
Once the potential threats are identified, the local component
solves a two-phase trajectory optimization problem. Note how-
ever that if a potential threat does not exist, the local component
solves a one-phase trajectory optimization problem. The aircraft
controller generates input commands to follow the optimal trajec-
tory, and applies them to the vehicle flight dynamics. The aircraft
controller loop monitors vehicle states and updates control inputs
with feedback rate fa . The guidance and navigation loop updates
the optimal trajectory with feedback rate fg given aircraft states
and the current global path approximation.
2.3. Global trajectory optimization
2.3.1. Obstacle resolution
The first step in the identification of an optimal global trajectory
is the resolution of multiple obstacles ahead of the aircraft’s current
flight path. This function entails a suitable characterization of
obstacles based on their location relative to the aircraft and each
other, as detected by the on-board sensor. We adopt a general sen-
sor model that captures the fundamental mechanism underlying
many such instruments. The model features an array of distance-
measuring rays originating at the sensor, uniformly structured
over the horizontal and vertical field of view. We assume the field
of view is symmetric about the x-axis in the vehicle coordinate
system GV and forward looking. The distance to an external body
along any ray is known whenever the ray intersects a surface of that
body, unless it exceeds a detection range specified in the model.
The relative angular placement of all sensor rays within the field of
view is prescribed so that the angle of each the ray in the vehicle
coordinate system is known. Accordingly, with measured distances
along known angles in GV the sensor data array provides the
location of all detection points. Sensor capabilities may be defined
by adjusting the vertical and horizontal field of view, density of
distance-measuring rays, and the detection range. We assume that
instrument error and signal return time are negligible, so that
the model is deterministic and instantaneous. We also assume
the detection range is isometric so that no variation exists along
vertical or horizontal field of view directions.
Obstacle resolution is hence realized as an operation on the
sensor data array where separate objects must be characterized
and distinguished. For the obstacle resolution, we consider two
clustering techniques: DBSCAN and spectral clustering technique.
In DBSCAN, the number of clusters is determined by the clustering
process itself, driven by the grouping of points that meet a user-
defined minimum density. The latter is established by two param-
eters that define the ϵ -neighborhood of a point p: the minimum
number of points minPts that must be within a distance ϵ from
point p. These required user-defined algorithm parameters are
lower-level data attributes that are more intuitive, meaningful,
and practical than emergent clustering characteristics, such as
number of clusters. Point density also provides an ideal mechanism
to resolve adjacent clusters while guaranteeing that a minimum
safe distance between them exists. Specifically for the current
application of obstacle resolution, ϵ can be used to explicitly set
the minimum distance between obstacles beyond which they are
resolved as separate objects.
In the spectral clustering algorithm, the number of clusters k
is initially identified through k-block diagonality of the Laplacian
matrix L. Once k is known, clusters are resolved with any algorithm
requiring k as an input, such as k-means. An adjacency matrix is
binary and denotes pairwise connectivity of nodes with a ‘‘1’’, and
‘‘0’’ otherwise. We introduce a characteristic distance d as the basis
of distance-based adjacency and define elements in A as follows:
Aij = 1,
 
if xi − xj  ≤ d
(1)
Aij = 0, if xi − xj  > d
 
where the vectors xi and xj denote data points, or vertices in the
graph-theoretic sense.
The overall approach is summarized in Algorithm 1. Caution
should be exercised whenever noise is present in the data because
it degrades k-block diagonality [28]. For applications where data
features high signal-to-noise ratio, or in the extreme case of com-
pletely deterministic data such as that produced by our determin-
istic sensor model, concerns associated with k-block diagonality
may be ignored.
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173 161

Fig. 1. Concept formulation and flow of the two-layer collision avoidance algorithm.

Fig. 2. New path planning architecture using a machine learning algorithm.

ways. First, it expedites the runtime of the clustering function.

We denote the distance Do as the user-defined minimum ob-
served distance between obstacles for resolution via sensor data
clustering. Accordingly, in our implementation of the two clus-
tering approaches discussed here, the characteristic distance pa-
rameter, ϵ for DBSCAN and d for spectral clustering with k-block
diagonality, are set to guarantee that sensed obstacles that appear
separated by a distance Do or more are resolved as separate objects.
Each of the clustering algorithms is implemented in a two-
dimensional domain where the sensor data has been collapsed
onto the yz plane of the vehicle coordinate system. Eliminating
obstacle depth affords significant computational efficiency in two
Second, the need to resolve obstacles based on depth (along the
longitudinal x axis of the vehicle coordinate system), and the
need to incorporate this information into a more complex global
trajectory optimization, is circumvented altogether.
On the other hand, exclusion of depth information can poten-
tially degrade the ability to solve for optimal avoidance trajectories
under certain conditions. Specifically, without depth information,
obstacles that are partially or totally blocked by nearer obstacles
are not resolved as separate objects. As a result, optimal trajec-
tories that may exist such that aircraft can navigate between up-
stream obstacles and partially blocked obstacles downstream are
not realized. This condition is examined in depth with a numerical
162 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
Algorithm 1 Spectral clustering
Inputs: point cloud information P ∈ Rn×d , distance constraint
between obstacles Do
(1) Distance-based adjacency n×n
∑ matrix Aij ∈ R , d = Do
(2) Degree matrix Dii = A ij if i = j, otherwise Dij = 0
j
(3) Unnormalized Laplacian matrix L = D − A
(4) Solve for k with k-block diagonality: rank(L) = n - k
(5) Treat each row of L as a point in Rk , and cluster into k clusters
via k-means
Outputs: Clustering result C1 , C2 , ... , Ck
Fig. 3. Identification of a potential threat.
simulation of a representative scenario in Section 3.3, where we
note that the performance of the proposed two-layer approach is
reduced to that of the single-layer alternative.
2.3.2. Identification of a potential threat
The second step in global trajectory optimization is to construct
global path alternatives commensurate with the obstacle informa-
tion produced in the previous step. We note however that there
exists an infinite number of feasible trajectories that maintain a
minimum safe distance from all obstacles. We proposed to iden-
tify potential threats in the global optimization loop, and solve
the two-phase trajectory optimization problem using the threat
information in the local trajectory optimization. Note that if the
potential threat does not exist, the local trajectory optimization
solves the one-layer trajectory problem.
The potential threat is selected by individual cluster informa-
tion and the vehicle velocity vector v = [vx vy vz ]. More specif-
ically, we project each point cluster onto a plane for which the
normal vector is parallel to x-direction (vx ) of the velocity vector,
and located at a distance corresponding to the nearest point of the
cluster along the x-direction of the navigation coordinate system.
It then projects the velocity vector onto individual planes. For
instance, in the case of Fig. 3, we project the velocity vector onto
three planes (plane 1, plane 2, and plane 3). The projected velocity
vector judges a potential threat. In other words, if the projected
vector on a plane is inside of the plane boundary, that cluster is
identified as a potential threat. In the example, the cluster on plane
2 is a potential threat. In some cases, there may not be any potential
threats. This case solves as a one-phase optimal trajectory problem
in the local trajectory optimization.
2.4. Local trajectory optimization
2.4.1. General multi-phase optimal trajectory problem
In the online process, the local trajectory optimization solves
two types of problems, namely, one- and two-phase trajectory
optimization. When a potential threat is not detected, the local tra-
jectory optimization solves the one-phase trajectory optimization
problem to reach the target position xt . When a potential treat is
detected, the local trajectory optimization solves the two-phase
trajectory optimization problem to avoid the obstacle resolved
with cluster C̃ and then reach the target position xt . We recognize
that obstacle avoidance and recovery towards a prescribed target
encompass the two fundamental phases of collision avoidance, and
that a multi-phase approach to trajectory optimization is therefore
well-suited for this problem. In this subsection, we only illustrate
the multi-phase framework since the one-phase framework is
relatively simple and can be easily defined from the multi-phase
framework.
In general, the multi-phase problem divides the trajectory into
n phases or segments (p ∈ 0, 1, 2, . . . , n) and sequentially solves
for the optimal trajectory in each phase p. The formulation is pred-
icated on the definition of a performance index or cost function,
dynamic constraints, path constraints, event constraints, and link
constraints. The cost function can be written as
n
∑ (p) (p)
min J(x̄(p) , u(p) , t) = 8(p) (x̄f , tf )
p=1
n ∫ (p)
∑ tf
+ L(p) (x̄(p) , u(p) , t)dt
(p)
p=1 t0
p
where the superscript denotes the pth phase, x̄ is a state vector, u
is a control input vector, t is time, t0 is initial time, and tf is terminal
time. 8 and L are terminal and transient costs, also called Mayer
and Lagrange costs in the optimal control theory context. Inclusion
of transient and terminal costs represents the most general for-
mulation of the cost function. However, for subclasses of optimal
trajectory problems, only transient or terminal costs will suffice.
Vehicle state equations are captured as dynamic constraints,
typically expressed in the form:
dx̄(p)
= f (p) (x̄(p) , u(p) , t). (3)
dt
Path constraints are algebraic inequalities that capture vehicle
flight performance limitations such as velocity, acceleration, and
thrust limits. These constraints can be represented as follows:
(p)
cmin ≤ c(p) (x̄(p) , u(p) , t) ≤ c(p)
max . (4)
Event constraints establish the conditions that must be satisfied
at the final time tf of each phase p, and are generally written in the
form
(p)
Emin ≤ E(p) (x̄(p) , u(p) , t) ≤ E(p)
max . (5)
Phase link constraints ensure continuous state transition be-
tween phases. For the link s between phases p − 1 and p, the
constraints are expressed as:
χ(s) (x̄f
(p−1)
, x̄(p) (p−1)
(tf ) − x̄(p) (t0 ) = 0, (p = s + 1).
0 ) = x̄ (6)
2.4.2. Two-phase trajectory optimization framework
To formulate the local-trajectory optimization, we suggest a
two-phase approach based on the multi-phase optimal trajectory
framework. The first phase (p = 1) is an avoidance phase, and
the second phase (p = 2) is a recovery phase. The cost function
and constraints are defined accordingly. Trajectory solutions can
be time-optimal, effort- or energy-optimal, or hybrid by combin-
ing the two. Other performance index formulations linked to the
vehicle dynamics are certainly possible, although time and energy
are very commonly used as the basis. The time-based approach
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173 163

minimizes trajectory duration, say in a hostile operating environ- are defined as

ment, and has been employed in a number of studies [22,23,31]. 1
⎡ ⎤
− 0 0
However, minimal-time trajectories impose high energy require- ⎢ τx ⎥
⎢ 1 ⎥
ments associated with aggressive maneuvers and accelerations, Ã = ⎢ 0 −
⎢ 0 ⎥,
⎥
which in turn strain energy management requirements and can ⎢ τy ⎥
⎣ 1⎦
degrade mission feasibility. On the other hand, a minimal-energy 0 0 −
τz
solution emphasizes mission energy management and efficiency,
1
⎡ ⎤
and is not driven by time or duration limitations. For this study, 0 0
⎢ τx ⎥ (9)
we adopt an energy-based performance index to assess effort- ⎢0 1 0⎥
⎢ ⎥
B̃ = ⎢ ⎥,
efficiency improvements attainable with the proposed collision ⎢ τy ⎥
avoidance concept. The performance index is defined as:
⎣ 1⎦
0 0
τz
n=2 ∫ (p)
tf
1∑
⎡ ⎤
0 ··· 0
J = u(p)T W(p) u(p) dt (7)
2 ⎢. .. ⎥ .
0̃n×n = ⎣ ..
(p)
t0
p=1 .⎦
0 ··· 0
where phase p (here limited to p = 1, 2) is denoted by the super-
script p , t0 and tf are initial and final times, respectively, u(p) ∈ R3 is τx , τy , and τz are time constants with respect to the three accelera-
(p) (p) (p)
(p) (p) (p)
the acceleration command input vector [ucx ucy ucz ]T , and W(p) tion state variables [ax ay az ]T .
Event constraints in the first phase prescribe the terminal vehi-
is a 3 × 3 weighting matrix presumed to be the same in both phases.
cle states that guarantee obstacle avoidance, and they are therefore
The 1/2 factor is inherent in the transient cost function L and is
fundamental for the entire collision avoidance trajectory problem.
simply placed outside the summation in Eq. (7) above. The cost In our approach, prescription of this terminal state is predicated on
function adopted here is based on the notion that work expended two conditions. First, the aircraft is in unaccelerated, level flight.
follows the commanded acceleration, squared to ensure positive This flight condition has been commonly used in the past, for
quantities, and has been used frequently in prior optimal trajectory instance by Moon and Kang [22]. It offers a simple and tractable
state that is justifiable in most cases when the target point after
problems (see for instance Ref. [32]). In the present formulation,
successful collision avoidance prescribes an altitude and velocity
the effort associated with the trajectory is entirely captured in the
comparable to that of the initial condition. We adopt this condition
transient cost expression, and a terminal cost component is not but note that it is simply suggested here, and that other conditions
necessary. may also be considered. Second, to enforce a safe distance from
In general, constraint complexity has a significant effect on any avoided obstacle, we introduce the variable (rs ). Three linear
the runtime performance of the online obstacle avoidance algo- inequality constraints are defined at a distance rs around the ob-
stacle cluster C̃ , one to the left, one to the right, and one above, as
rithm, so the choice of constraints and their inherent sophisti-
shown in Fig. 4.
cation must be considered carefully. This is particularly relevant
Accordingly, we define the terminal state as
for dynamic constraints where simplified flight dynamics can be
(1)
adopted whenever possible instead of complex non-linear alter- tf = free
natives. Moon and Kang [22,23] propose first-order acceleration x(1) (tf ) = [x(1) (tf ) free free]T (x(1) (tf ) = min C̃x − rs )
dynamic equations for simplified kinematics of an optimal tra- (10)
v(1) (tf ) = [u(1) (tf ) 0 0 ]T
jectory problem. We employ this first-order approximation for
dynamic constraints to achieve a computationally and dynamically a(1) (tf ) = [0 0 0]T
feasible solution. This choice for low order dynamics constraints
where C̃x is the x position information of C̃ . Note that the safe
is a practical one, driven by the present focus to demonstrate the position along the x axis, x(1) (tf ), is prescribed according to rs too.
proposed two-layer approach in relation to the classical one-layer The y and z positions are free variables that are only restricted by
alternative. An implementation of our method with higher order the following inequality constraints, pursuant to the safe distance
dynamics is immediately feasible, but increased runtime should be from the obstacle for the avoidance trajectory:
⎧
expected as a result. ⎪ y ≤ C̃ymin + Mt1
⎨
The simplified kinematic constraints are as follows: y ≥ C̃ymax − Mt2
⎪
⎩
z ≥ C̃zmax − Mt3 (11)
⎡ ⎤ ⎡ ⎤⎡ ⎤
ẋ(p) 0̃3×3 Ĩ3×3 0̃3×3 x(p) [ ]
0̃6×3
⎣v̇(p) ⎦ = ⎣0̃6×6 Ĩ3×3 ⎦ ⎣ v(p) ⎦
+ u(p) (8) t = [t1 t2 t3 ], ti ∈ [1, 0]
B̃
ȧ(p) Ã a (p)
M is a large number
(p) (p) (p) (p) T (p)
where x is the position vector [x y z ] , v is the where C̃ymin and C̃ymax are the minimum and maximum y positions,
velocity vector [u(p) v (p) w (p) ]T , a is the acceleration vec- respectively, and C̃zmax is the maximum z position, in the point
(p) (p) (p)
tor [ax ay az ]T , and u(p) is the acceleration command cloud C̃ . The binary vector t identifies which of the three obstacle
(p) (p) (p) T
[ucx ucy ucz ] . These states are expressed in a navigation coor- avoidance constraints is active with vector element ti = 0, and
dinate system. Ĩ3×3 is a 3 × 3 identity matrix, and Ã, B̃, and 0̃n×m tj̸=i = 1 otherwise. Identification of the active constraint, and
164 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173

stated as simple load factor constraints:

|ax | < gxmax
⏐ay ⏐ < gymax
⏐ ⏐
(15)

|az + g | < gzmax

where gxmax , gymax , and gzmax are maximum allowable accelera-
tions. We also adopt a minimum altitude constraint as a path
constraint to guarantee flight above some operational minimum
to prevent possible collisions with other ground assets.

z ≥ zmin . (16)

The last constraint is the linkage constraint at the phase transition

point to enforce state continuity across phases:

x̄(p−1) (tf ) = x̄(p) (t0 ). (17)

Fig. 4. Notional depiction of the three obstacle avoidance trajectory event con-
Since the proposed framework of the optimal collision avoidance
straints at a safe distance rs around the obstacle cluster C̃ .
problem has two phases, in the phase linkage constraint (Eq. (17)),
p is two. The formulated framework of the multi-phase optimal
trajectory problem is solved by the Gauss Pseudospectral method
(GPM) provided by the open-source software GPOPS. This Gauss
therefore definition of the binary vector t, is solved for as follows: Pseudospectral technique was developed by Benson [33,34], and
⏐ ⏐ ⏐ ⏐ it was advanced and validated by empirical cases studies from
D = [⏐(min C̃y − rs ) − yp ⏐ , ⏐(max C̃y + rs ) − yp ⏐ ,
⏐ ⏐ ⏐ ⏐
Huntinton et al. [35–37]. This GPM technique employs an orthogo-
⏐ ⏐ nal collocation method based on the Legendre–Gauss points. The
⏐(max C̃z + rs ) − zp ⏐] GPOPS software provides a MATLAB interface with a non-linear
⏐ ⏐
⏐ ⏐ programming problem solver, SNOPT [38,39].
t = [0 1 1], if min D = ⏐(min C̃y − rs ) − yp ⏐
⏐ ⏐
(12) We further examine the proposed two-layer algorithm in terms
⏐ ⏐ of the rotation angle of the cuboid obstacle relative to the initial
t = [1 0 1], if min D = ⏐(max C̃y + rs ) − yp ⏐
⏐ ⏐ navigation coordinate system. The impact of this geometric feature
is assessed by a comparative assessment of a representative sce-
⏐ ⏐
nario where a cuboid obstacle faces the aircraft navigation coordi-
t = [1 1 0], if min D = ⏐(max C̃z + rs ) − zp ⏐ .
⏐ ⏐
nate system at the start of the simulation, and a second case where
the same object is rotated. The time sequence of the avoidance
Here, yp and zp are the vehicle position in y and z projected with
(1) trajectories that result from the simulation, as well as the vehicle
the current velocity vector v0 (t) onto a plane perpendicular to the response, are captured in Figs. 5 and 6. In both figures, the yellow
velocity vector at a downrange distance x(1) (tf ). lines are the sensor rays, the black line is the vehicle response
Event constraints for the second phase dictate the terminal state trajectory, and the blue line is the result of the collision avoidance
at the target position, for which we recommend unaccelerated trajectory at a specific time t. The results show that in the first
level flight, as follows: case where the obstacle faces the aircraft, the resulting avoidance
trajectory is generated to the right of the obstacle. For the case
(2) where the obstacle has been rotated, the energy optimal trajectory
tf = free
]T is to the left of the obstacle, which is indicative that the projected
x(2) (tf ) = x(2) (tf ) y(2) (tf ) z (2) (tf )
[
velocity vector is closer to the left edge of the sensor data point
(13) projection of the rotated building. In both cases, the solutions are
v(2) (tf ) = [vx(2) (tf ) 0 0] found to be energy optimal, and more generally demonstrate that
a(2) (tf ) = [0 0 0]. alignment of the cuboid obstacle faces to the navigation coordinate
system is not required to attain energy-optimal trajectories.
Path and link constraints are defined equally for the avoidance and
recovery phases. Path constraints curtail vehicle dynamic perfor- 3. Numerical simulation
mance and include limits on velocity and acceleration, keeping the
3.1. Simulation of unmanned aircraft dynamics, controller, and sensor
vehicle within the flight envelope to prevent unfeasible maneu-
vers. Path constraints are defined as follows:
For the numerical simulation, we assume a small electric fixed-
umin ≤ u ≤ umax wing unmanned aircraft, and we adopt the Aerosonde specification
taken directly from the literature [40,41]. Table 1 summarizes
vmin ≤ v ≤ vmax airframe parameters, along with those assumed for the airborne
(14)
wmin ≤ w ≤ wmax sensor and collision avoidance scheme. For simplicity, the vehicle
√ equations of motion assume a point mass model, as reported in
Vmin ≤ u2 + v 2 + w 2 ≤ Vmax previous work [42–44]. For the aircraft tracking controller, we em-
ploy the standard feedback control structure. The on-board sensor
where umin , vmin , and wmin are the low speed limits, and umax , vmax , is modeled as a generic light detection and ranging instrument, or
and wmax are the high speed limits. The acceleration constraints are LiDAR, with no instrument uncertainties.
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173 165

Fig. 5. Time sequence of the optimal trajectory when an obstacle is aligned with the initial navigation frame.

Fig. 6. Time sequence of the optimal trajectory when an obstacle is not aligned with the initial navigation frame.

3.2. Comparative analysis of clustering algorithms for obstacle reso-
lution
We compare the DBSCAN and spectral clustering algorithms
for the present application in terms of prediction success rate and
computational time. These figures of merit are evaluated over a
design of experiments based on the setup illustrated in Fig. 7,
where two identical, generic, cuboid obstacles lay ahead of the
aircraft flight path. The x, y, and z axes shown are the vehicle
coordinate system GV . The sensor data collection is assumed to
be instantaneous, and the clustering algorithms executed at that
instant only, so that there is no trajectory optimization or flight
simulation, only clustering evaluation at the instant depicted. The
distance 2D between obstacles, referenced to the origin of the y
axis, is sampled as follows: D = [20 40 60 80 100]. Similarly,
the down-range distance between the unmanned aircraft and each
of the two obstacles is varied with the x axis location of the obstacle
centroid as follows: L1, L2 = [1100 1200 1300 1400 1500].
We combine the three independent factor variations into a full
factorial design of experiments with 125 samples. The UAV sensor
parameters and the minimum separation distance between obsta-
cles, Do , are described in Table 1. The parameter Do is mapped to
the characteristic distance in the clustering algorithm of choice, so
that obstacles are resolved as separate objects only if the distance
between them is greater than Do . Note that to allow safe flight
between resolved obstacles, the distance between them must be at
least twice the minimum separation distance between the aircraft
and any obstacle, rs .
Do ≥ 2rs . (18)
166 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173

Table 1
Description of UAV parameters.
UAV parameters Variable Value Unit
Weight w 29.76 [lb]
Planform area S 6.1 [ft2 ]
Area swept out by the propeller Sp 0.1348 [ft2 ]
Propeller aerodynamic coefficient Cp 1
Efficiency constant of the motor Km 8
Vehicle parameters
Lift coefficient at zero angle of attack CL0 0.28
Lift curve slope CLα 3.45
Aspect ratio AR 10.7
Span efficiency e 0.9
Zero-lift drag coefficient CD0 0.03

Distance range Sr 1500 [ft]

Azimuth range AZ 60 [deg]
Sensor parameters
Elevation range EL 45 [deg]
Sensor resolution θsen 1.5 [deg]

Minimum distance constraint between obstacles Do 150 [ft]

Minimum separation distance between an obstacle and a vehicle rs 70 [ft]
Collision avoidance parameters
Minimum altitude zmin 200 [ft]
Maximum acceleration gmax 64.34 [ft/s2 ]

Update rate of guidance and navigation loop fg 1 [Hz]

Other parameters
Update rate of aircraft control loop fa 10 [Hz]

Figs. 8–11 illustrate four representative cases in the evaluation of

clustering algorithms. Based on the results obtained, we opt for
DBSCAN as the preferred clustering algorithm in the global-path
optimization.
The clustering algorithm has a computational burden on the
order of 0.1 [s]. Next, the global optimization takes, on average,
0.019 [s]. If an obstacle is found to present a conflict for the current
trajectory, the two-phase local trajectory optimization takes, on
average, 3.5 [s], and if no obstacle threat exists then the single
phase local optimization takes approximately 0.49 [s]. This means
that for cases where obstacle conflicts are found, the relative cost
of the clustering algorithm is marginal, on the order of 3.4%. On
the other hand, for cases where there are no obstacle threats, the
computational burden of the clustering algorithm is 24.3%.
(a) Top view.
3.3. Assessment of two-layer collision avoidance with numerical sim-
ulation

To test the hypothesis that our two-layer collision avoid-

(b) Side view.
Fig. 7. Simulated experiment setup for comparative analysis of clustering algo-
rithms. (Aircraft not shown to scale).
This experiment was conducted on a commercially available
desktop computer with a 3.40 GHz Intel(R) Core(TM) i7-2600 pro-
cessor and 8 GB of RAM. Results of the clustering experiments are
summarized in Table 2 and show that both algorithms accurately
resolve clusters with a 100% success rate. That is, when the gap
between obstacles is larger than Do (i.e., 2D ≥ Do ), both algorithms
exactly resolve two clusters regardless of variation of the distances
L1 and L2 .
In terms of computational efficiency, DBSCAN is faster and fea-
tures lower variance across experiments. We attribute the greater
runtime mean and variance of spectral clustering in part to the use
of k-means with random initialization for the initial center points.
ance algorithm yields more energy efficient trajectories without
incurring a prohibitive computational burden relative to the
single-layer approach, we conduct an experiment with numerical
simulations. The experiment is comprised of three use cases de-
signed to be relevant and representative for UAS applications of
interest, and to capture critical arrangements of relative obstacle
placement in terms of on-board sensing and optimal avoidance
trajectory execution. Simulations for the four use cases are run
with both methods, single- and two-layer, using the same initial
conditions, set to steady level flight at 500 feet altitude with a
velocity of 70 feet/s.
The first use case is illustrated in Fig. 12 with the trajectory
simulation results of the two collision avoidance methods. The
obstacle-free trajectory between the initial and final points is
shown in red, and the executed collision avoidance trajectory
with each method is shown in dotted blue. The first scenario
features two obstacles of comparable altitude (∼1000 feet) and
footprint, located at the same downrange distance (∼3800 feet),
and separated by a lateral distance (∼175 feet) greater than twice
the safe distance defined for aircraft-to-obstacle separation. This
initial scenario is intended to characterize the benefits of safe flight
between adjacent obstacles. Results show that with the one-layer
obstacle avoidance algorithm the aircraft maneuvers left around
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
Table 2
Comparison assessment results of the clustering algorithms.
DBSCAN
Prediction success rate (%)
Average 100
Standard deviation 0 0.027
(a) Raw data.
(b) Clustering results.
Fig. 8. Clustering results (D = 20 ft, L1 = 1100 ft, L2 = 1100 ft).
the left obstacle, since that route is more energy efficient than
flying over the two obstacles or around the right obstacle. As
expected, in this simulation, the aircraft treats both obstacles as
a single object and executes the most energy-efficient trajectory
to avoid it and meet the final point. In contrast, the result of
two-layer obstacle avoidance algorithm shows that the aircraft
flies a trajectory through the gap between the two obstacles, with
only a minor course correction too small to be appreciated in the
figure. Again, the result is reasonably expected as the gap between
the two obstacles exceeds the minimum separation, allowing the
global optimizer to resolve them as two separate objects, and
for the local optimizer to produce an energy-optimal avoidance
trajectory between the two. Fig. 13 shows the time history of
velocity, acceleration, and separation distance between the aircraft
and the obstacles. The velocity responses show that the single-
layer solution features greater variation than that for the proposed
two-layer structure, in this particular case due to the maneuver
around the left obstacle. Hence, unlike the one-layer structure, the
acceleration of the two-layer structure remains approximately at
zero, which is indicative of a more energy efficient trajectory. The
response of the separation distance shows that both techniques
generate a safe avoidance route.
The second use case, depicted in Fig. 14, builds upon the first
one and is designed to test cases when there is a third obstacle
immediately behind the gap between the first two obstacles. As
before, the one-layer algorithm avoids the first two obstacles by
167
Spectral method
Computational time (s) Prediction success rate (%) Computational time (s)
0.101 100 0.133
0 0.214
(a) Raw data.
(b) Clustering results.
Fig. 9. Clustering results (D = 40 ft, L1 = 1100 ft, L2 = 1500 ft).
maneuvering to the left of the left obstacle, and then keeping to
the left of the third obstacle in the back. The two-layer algorithm
executes an avoidance trajectory between the first two obstacles,
and then around the right of the third obstacle behind them. Fig. 15
presents the time history of velocity, acceleration, and the separa-
tion distance. As with the previous use case, the time responses
indicate that the two-layer approach features smaller velocity
variations, less energy required, and comparable separation to
obstacles.
The third use case, depicted in Fig. 16, reverses the obstacle
arrangement, presenting first a single obstacle whose footprint
and orientation are intentionally chosen to significantly block the
two obstacles behind it. This scenario is intended to test the case
where information about downrange obstacles cannot be acquired
until a closer obstacle is cleared. Moreover, it tests the feasibility
and efficacy of the proposed method with regard to our choice
to cluster sensor data in a 2D domain, that is, collapsed to the
xy plane, so that obstacles are not resolved based on depth data.
Results shows that the two obstacle avoidance algorithms have
the same initial obstacle avoidance trajectory to avoid the first
obstacle. After avoiding the first obstacle the aircraft detects two
downstream obstacles and from that point on the two algorithms
produce different avoidance trajectory solutions. In the one-layer
obstacle avoidance trajectory, the aircraft flies above the two ob-
stacles because it computes the energy-efficient avoidance trajec-
tory resolving the two obstacles as a single object. The two-layer
168 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173

(a) Raw data.

(a) One-layer obstacle avoidance algorithm.

(b) Clustering results.

Fig. 10. Clustering results (D = 100 ft, L1 = 1500 ft, L2 = 1100 ft).

(b) Two-layer obstacle avoidance algorithm.

Fig. 12. Numerical simulation results of obstacle avoidance algorithms in the first
scenario.

(a) Raw data.

(b) Clustering results.
Fig. 11. Clustering results (D = 100 ft, L1 = 1500 ft, L2 = 1500 ft).
algorithm resolves two separate obstacles and executes an energy
optimal trajectory through the gap between them. Fig. 17 is the
time response. The time responses show that the velocity profile is
exactly the same before the aircraft detects the two downstream
obstacles, with notable differences in velocity variations after that.
The one-layer acceleration response is consistent with greater
energy demands.
The fourth case, depicted in Fig. 18, includes three obstacles
intentionally arranged so that view of the two downstream obsta-
cles is partially and totally blocked, respectively, by the upstream
obstacle. Like the third scenario discussed above, this scenario is
representative of conditions where the choice to exclude obstacle
depth information can be characterized in terms of impacts on
energy requirements for avoidance trajectories. The result shows
that both algorithms present the same avoidance trajectories be-
cause the clustering algorithm in the global optimization does
not resolve the two obstacles and instead treats them as a single
obstacle. Fig. 19 is the time response. Both algorithms generate
exactly the same response for velocity, acceleration, and separation
distance.
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173 169

(a) One-layer obstacle avoidance algorithm.

Fig. 13. Time response of velocity, acceleration and minimum distance between the
aircraft and obstacles in the first scenario.

We quantitatively assess the performance of both algorithms

for the four scenarios by estimating the work done by the aircraft,
or conversely, the energy required to execute each trajectory.
We adopt the classical energy formulation for flight performance,
where the work performed by the aircraft, i.e. the energy ex-
pended, is estimated as thrust (force) applied over distance, where
it is assumed that the thrust and velocity vectors are aligned so
that thrust is always in the direction of motion. In this formulation,
there is no work associated with lift as it is always perpendic- (b) Two-layer obstacle avoidance algorithm.
ular to the trajectory. The work associated with drag is always
an energy loss. Weight work corresponds to the change in po- Fig. 14. Numerical simulation results of obstacle avoidance algorithms in the
tential energy. The work associated with the imbalance between second scenario.
thrust, drag, and the component of weight aligned with them,
resulting in longitudinal accelerations, corresponds to the change
in kinetic energy. The latter is often adjusted via thrust input.
Accordingly, executing a trajectory is given by
T = D + W sin γ + F ∫
WT = Tds, (20)
T ds = Dds + W ds sin γ + F ds c
(19)
WT = WD + WW + WF where T is thrust, c is an avoidance trajectory, and ds is the length
of a discrete segment along the avoidance trajectory.
WT = WD + 1PE + 1KE . Table 3 summarizes the measurement results of each method
Here, γ is the flight path angle, F is the thrust-controlled force im- under the four different scenarios. Results clearly indicate that the
balance between thrust, drag, and the weight component aligned proposed method requires less energy than the one-layer collision
to the flight path, ds is an infinitesimal segment along the flight avoidance method. It is worth noting too that both algorithms
trajectory, and WT denotes the work associated with force T , as establish energy-optimal trajectories around resolved obstacles, so
is also the case for forces D, W , and F . Note that for cases where
that all energy-efficiency benefits observed can only be attributed
γ = 0, the work associated with weight is zero, and there is no
to the resolution of separate objects from the proposed two-layer
change in potential energy, so 1PE = 0. Similarly, for the case of
null imbalance force, i.e. F = 0, the work associated with F is zero, approach.
and there is no change in kinetic energy, so 1KE = 0. Results clearly show that the proposed two-layer approach
Recognizing that thrust, and the work it performs over ds, generates more energy-efficient obstacle avoidance trajectories.
equate to changes in kinetic and potential energy while over- The inclusion of the global trajectory layer results in additional
coming drag, then the work performed (or energy expended) in computational expense, where the most demanding step is the
170 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173

Table 3
Summary of numerical simulation.
Scenario One-layer structure (lb ft) Two-layer structure (lb ft) Energy difference (lb ft)
W1 W2 W1 − W2
1 19 816 18 690 1125
2 19 816 18 896 920
3 23 353 21 645 1708
4 18 773 18 773 0

(a) One-layer obstacle avoidance algorithm.

Fig. 15. Time response of velocity, acceleration and minimum distance between the
aircraft and obstacles in the second scenario.

solution of the clustering problem. Based on our runtime assess-

ments of DBSCAN presented in Table 2, we find that this cost
is acceptably low in comparison to the typical runtime of the
local trajectory optimization. Additional runtime benefits may be
attained via programming improvements or parallel computing
techniques.

3.4. Simulation in a realistic urban environment (b) Two-layer obstacle avoidance algorithm.

In this section, we demonstrate the proposed two-layer obsta-

cle avoidance algorithm in a representative urban environment.
In order to generate a fully realistic environment, we generate a
representation of a section of downtown San Diego, as shown in
Fig. 20(a). To generate this model of a real urban setting, we use
the actual geometric information from publicly available LiDAR
surveys. The result of the urban model is presented in Fig. 20(b).
In this model, buildings are represented as cuboid bodies with a
minimum altitude of 200 feet.
We identify the ten trajectories traversing this environment
with the most conflicts with obstacles, as illustrated in Fig. 21 with
the initial and terminal waypoints. To identify these challenging
routes and to avoid trivial cases with no obstacle conflicts, we
define five initial and terminal conditions along the x and y axis
and compute the initial optimal trajectories. Then, we select the ten
routes with the greatest number of conflicts. Note that the initial
Fig. 16. Numerical simulation results of obstacle avoidance algorithms in the third
scenario.
altitude is assumed to be 350 feet, and the initial flight condition is
a level flight at a speed of 70 feet/s.
Table 4 summarizes the energy quantification results of the
ten avoidance trajectories. The results clearly illustrate that the
proposed two-layer algorithm yields more energy efficient trajec-
tories, with the exception of the sixth case, where the optimal
trajectory is exactly the same as that of the single-layer method.
The average improvement is approximately 7.1%. Two representa-
tive simulation results, shown in Figs. 22 and 23, indicate typical
differences observed, specifically where the traditional one-layer
approach yields trajectories that fly over the all obstacles in a dense
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173 171

Table 4
Energy quantification of the urban scenarios.
Scenario no. 1 2 3 4 5 6 7 8 9 10
One layer [lb ft] 19746.39 18074.16 17997.87 16962.03 17131.05 15879.09 19681.15 16095.72 19713.92 18725.62
Two layer [lb ft] 16777.26 16436.40 17178.83 16137.37 16381.19 15879.09 17161.34 15644.07 17371.21 17621.44
Improvement [%] 15.04 9.06 4.55 4.86 4.38 0.00 12.80 2.81 11.88 5.90

Fig. 19. Time response of velocity, acceleration and minimum distance between the
Fig. 17. Time response of velocity, acceleration and minimum distance between the aircraft and obstacles in the fourth scenario.
aircraft and obstacles in the third scenario.

4. Conclusion

Many existing collision avoidance concepts using on-board

Fig. 18. Numerical simulation results of obstacle avoidance algorithms (one-layer,
and second-layer) in the fourth scenario.
region, whereas the proposed two-layer approach yields trajec-
tories where the aircraft flies through the gap between obstacles
whenever sufficient distance exists between them.
sensor data fail to generate efficient trajectories because, even
though multiple objects may be detected, the collision avoidance
algorithm treats them as a single conglomerate regardless of their
number and relative location. Resulting avoidance trajectories are
therefore optimal only in the context of a large avoidance region
and fail to exploit free space between obstacles that may allow
for more efficient solutions. The approach proposed in this pa-
per features a two-layer multi-phase optimal collision avoidance
algorithm that entails global- and local-path optimization. The
purpose of the global-path optimization is to resolve multiple
obstacles as separate objects, given that they satisfy a separation
minimum, with sensor data clustering. The global-path optimiza-
tion also identifies a potential threat using the cluster data of
resolved objects. We identify two candidate clustering techniques
that allow for the parametric specification of a cluster separation
minimum for obstacle resolution: DBSCAN and spectral clustering.
A design of simulation experiments is executed to evaluate both
techniques in terms of prediction success rate and computational
runtime. Although both methods are found to have 100% rates
for the simulations performed, DBSCAN featured lower runtime
mean and variance, thus emerging as the recommended clustering
technique for the proposed concept.
Numerical simulations with status-quo one-layer collision
avoidance and the proposed two-layer concept are conducted
to compare their performance in terms of the energy-efficiency
172 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
(a) San Diego downtown (Google Earth image).
(b) Simulation environment.
Fig. 20. Urban simulation environment (San Diego).
Fig. 21. Ten initial trajectories in the urban scenario.
of executed trajectories. Four relevant use cases are utilized to
examine multi-obstacle placements of interest, assess the feasi-
bility of the proposed concept, and test the main hypothesis of
this work. Results indicate that trajectory solutions with the two-
layer avoidance method provide tangible energy efficiency benefits
while incurring a very small additional computational cost.
In order to further test, demonstrate, and validate the effec-
tiveness of the proposed algorithm, HIL (Hardware-In-the-Loop) or
unmanned flight tests are requisite, particularly since actual sensor
measurement data can contain significant noise and uncertainty
for which modifications to the method may be required. At this
point, however, we do not attempt to propose or speculate on what
Fig. 22. Numerical simulation result in the case with high energy difference (second
scenario).
Fig. 23. Numerical simulation result in the case with low energy difference (third
scenario).
those improvements may be. The actual flight tests and HIL tests to
address these factors remain as future work.
Collision avoidance with on-board sensors is a relatively new
and quickly evolving field where many areas of additional work are
warranted. Regarding the work here reported, we argue in favor of
a broader benchmarking study where a large set of multi-obstacle
scenarios is utilized, potentially with hardware-in-the-loop real-
time simulations. Moreover, this two-layer collision avoidance al-
gorithm has a flexible structure. In other words, the local trajectory
optimization can be realized with any of a myriad of obstacle
avoidance formulations.
We also believe that extension to a probabilistic study is a natu-
ral next step so that uncertainties associated with instrumentation,
on-board computer runtime, and flight dynamics and controls
governing trajectory execution, are treated explicitly. Ultimately,
the work presented here outlines a new concept that offers the
potential of significant benefits for the rapidly growing field of UAS
applications and the increasing need to develop safety-assuring
technologies.
Acknowledgment
The authors thank Dr. Jongki Moon for his insights and expert
input on this work.
References
[1] U. S. Department of Transportation, Air traffic organization policy (2014).
[2] A.R. Lacher, D.R. Maroney, A.D. Zeitlin, Unmanned aircraft collision avoidance–
technology assessment and evaluation methods, in: The 7th Air Traffic Man-
agement Research & Development Seminar. Barcelona, Spain (2007).
 Y. Choi et al. / Robotics and Autonomous Systems 98 (2017) 158–173
[3] K. Dalamagkidis, K.P. Valavanis, L.A. Piegl, On unmanned aircraft systems
issues, challenges and operational restrictions preventing integration into the
national airspace system, Progress in Aerospace Sciences 44 (7) (2008) 503–
519.
[4] S.M. La Valle, Planning Algorithms, Cambridge University Press, 2006.
[5] D. Levine, B. Luders, J.P. How, Information-rich path planning with general
constraints using rapidly-exploring random trees, in: AIAA Infotech Aerospace
Conference (2010).
[6] Y.-J. Ho, J.-S. Liu, Collision-free curvature-bounded smooth path planning
using composite bezier curve based on voronoi diagram, in: Computational
Intelligence in Robotics and Automation (CIRA), 2009 IEEE International Sym-
posium on. IEEE (2009), pp. 463–468.
[7] L. De Filippis, G. Guglieri, F. Quagliotti, Path planning strategies for UAVs in 3d
environments, J. Intell. Robot. Syst. 65 (1) (2012) 247–264.
[8] O. Khatib, Real-time obstacle avoidance for manipulators and mobile robots,
Int. J. Robot. Res. 5 (1) (1986) 90–98.
[9] G. Li, A. Yamashita, H. Asama, Y. Tamura, An efficient improved artificial
potential field based regression search method for robot path planning, in:
IEEE International Conference on Mechatronics and Automation (2012), pp.
1227–1232.
[10] D. Panagou, Motion planning and collision avoidance using navigation vector
fields, in: IEEE International Conference on Robotics and Automation (ICRA)
(2014), pp. 2513–2518.
[11] T.T. Mac, C. Copot, A. Hernandez, R.D. Keyser, Improved potential field method
for unknown obstacle avoidance using UAV in indoor environment, in: IEEE
14th International Symposium on Applied Machine Intelligence and Informat-
ics (SAMI) (2016).
[12] X.-Y. Li, L. Jie Guo, Constructing affinity matrix in spectral clustering based on
neighbor propagation, Neurocomputing 97 (2012) 125–130. Elsevier.
[13] J.-W. Park, H.-D. Oh, M.-J. Tahk, UAV collision avoidance based on geometric
approach, in: SICE Annual Conference, 20-22 August (2008).
[14] A. Mujumdar, Nonlinear geometric and differential geometric guidance of
UAVs for reactive collision avoidance. Tech. rep., DTIC Document (2009).
[15] A. Mujumdar, R. Padhi, Reactive collision avoidance of using nonlinear geo-
metric and differential geometric guidance, J. Guid. Control Dyn. 34 (1) (2011)
303–311.
[16] A. Chakravarthy, D. Ghose, Obstacle avoidance in a dynamic environment: A
collision cone approach, Systems, Man and Cybernetics, Part A: Systems and
Humans 28 (5) (1998) 562–574.
[17] Y. Watanabe, A.J. Calise, E.N. Johnson, Vision-based obstacle avoidance for
UAVs, in: AIAA Guidance, Navigation and Control Conference. Hilton Head,
South Carolina (2007).
[18] S. Tom, B.D. Moor, E. Feron, J. How, Mixed integer programming for multi-
vehicle path planning, in: European control conference. vol. 1. (2001), pp.
2603–2608.
[19] Y. Kuwata, J. How, Three dimensional receding horizon control for UAVs, in:
AIAA Guidance, Navigation, and Control Conference and Exhibit. Providence,
Rhode Island (2004).
[20] D. Mellinger, A. Kushleyev, V. Kumar, Mixed-integer quadratic program tra-
jectory generation for heterogeneous quadrotor teams, in: Robotics, (ICRA), A.
(Eds.), IEEE International Conference. (2012), pp. 477–483.
[21] G.N. Vanderplaats, Multidiscipline Design Optimization, Vanderplaats Re-
search & Development, Inc., 2007.
[22] J. Moon, J. Prasad, Minimum-time approach to obstacle avoidance constrained
by envelope protection for autonomous UAVs, in: Proceedings of AHS 65th
Annual Forum. Grapevine, TX (2009).
[23] K. Kang, J.V.R. Prasad, Development and flight test evaluations of an au-
tonomous obstacle avoidance system for a rotary-wing UAV, Unmanned Sys-
tems 1 (1) (2013) 3–19.
[24] Y. Peng, D. Qu, Y. Zhong, S. Xie, J. Luo, J. Gu, The obstacle detection and obstacle
avoidance algorithm based on 2-d lidar, in: 2015 IEEE International Conference
on Information and Automation (2015).
[25] D.T. Pham, S.S. Dimov, C.D. Nguyen, Selection of k in k-means clustering.
Proceedings of the Institution of Mechanical Engineers, Part C: Journal of
Mechanical Engineering Science 219 (1) (2005) 103–119.
[26] H. Jia, S. Ding, X. Xu, R. Nie, The latest research progress on spectral clustering,
Neural Comput. Appl. 24 (2014) 1477–1486.
[27] A.Y. Ng, M.I. Jordan, Y. Weiss, On spectral clustering: Analysis and an algorithm,
Advances in Neural Information Processing System 2 (2002) 849–856.
[28] J. Feng, Z. Lin, H. Xu, S. Yan, Robust subspace segmentation with block-diagonal
prior. (2014), pp. 3818–3825.
[29] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, A density-based algorithm for dis-
covering clusters in large spatial databases with noise. vol. 96, International
Conference on Knowledge Discovery and Data Mining (KDD-96) (1996).
173
[30] S. Maniatopoulos, D. Panagou, K.J. Kyriakopoulos, Model predictive control for
the navigation of a nonholonomic vehicle with field-of-view constraints, in:
IEEE (Ed.), American Control Conference (ACC) (2013), pp. 3967–3972.
[31] S. Sundar, Z. Shiller, Time-optimal obstacle avoidance, in: Proceedings of the
IEEE International Conference on Robotics and Automation. vol. 3. Nagoya,
Japan (2013).
[32] M. Anderson, J. Sverdrup, J. Lopez, J. Evers, A comparison of trajectory deter-
mination approaches for small uav’s, in: AIAA Atmospheric Flight Mechanics
Conference and Exhibit. Keystone, Colorado (2006).
[33] D. Benson, A gauss pseudospectral transcription for optimal control, Ph.D. The-
sis, Department of Aeronautics and Astronautics. Ph.D. Thesis, Massachusetts
Institute of Technology (2004).
[34] D.A. Benson, G.T. Huntington, T.P. Thorvaldsen, A.V. Rao, Direct trajectory
optimization and costate estimation via an orthogonal collocation method, J.
Guid. Control Dyn. 29 (6) (2006) 1435–1440.
[35] G.T. Huntington, D.A. Benson, A.V. Rao, Design of optimal tetrahedral space-
craft formations, Journal of the Astronautical Sciences 55 (2) (2007) 141–169.
[36] G.T. Huntington, D.A. Benson, J.P. How, C.L.D.N. Kanizay, A.V. Rao, Computation
of boundary controls using a gauss pseudospectral method, in: 2007 Astrody-
namics Specialist Conference. Mackinac Island, Michigan (2007).
[37] G.T. Huntington, A.V. Rao, Optimal reconfiguration of spacecraft formations
using the gauss pseudospectral method, J. Guid. Control Dyn. 31 (3) (2008)
689–698.
[38] P.E. Gill, W. Murray, M.A. Saunders, SNOPT: An SQP algorithm for large-scale
constrained optimization. Tech. rep., SIAM Review (2005).
[39] P.E. Gill, W. Murray, M.A. Saunders, User’S Guide for SNOPT Version 7: Software
for Large Scale Nonlinear Programming, Stanford University, 2006.
[40] F. Wieland, D. DeLaurentis, G. Kubat, Modeling and simulation for UAS in the
NAS. Tech. rep., NASA (2012).
[41] R.W. Beard, T.W. McLain, Small Unmanned Aircraft: Theory and Practice,
Princeton University Press, 2012.
[42] W. Lin, Distributed UAV formation control using differential game approach,
Aerosp. Sci. Technol. 35 (2014) 54–62.
[43] P.K. Menon, G.D. Sweriduk, B. Sridhar, Optimal strategies for free-flight air
traffic conflict resolution, J. Guid. Control Dyn. 22 (2) (1999) 202–211.
[44] Y. Choi, H. Jimenez, D.N. Mavris, Statistical gain-scheduling method for aircraft
flight simulation, Aerosp. Sci. Technol. 46 (2015) 493–505.
Youngjun Choi is research faculty at the School of
Guggenheim Aerospace Engineering, the Georgia Institute
of Technology. Dr. Choi received Ph.D. from that institution
in 2016. He was a visiting researcher at the Centre for
Power Transmission and Motion Control at the University
of Bath and worked at the Agency for Defense Develop-
ment. His research areas are guidance, navigation, and
control for autonomous vehicles, machine learning algo-
rithms, and aerospace system design.
Dr. Jimenez is research faculty at the Daniel Guggen-
heim School of Aerospace Engineering, Georgia Institute
of Technology. Dr. Jimenez received his Ph.D. from that
institution in 2009. He works at the Aerospace Systems
Design Laboratory where he serves as the chief of UAS-
Airspace Integration branch. In addition, he leads research
efforts in areas of systems analysis, general aviation safety,
and aviation environmental impacts.
Dimitri Mavris is S.P. Langley Regents Professor at the
Daniel Guggenheim School of Aerospace Engineering,
Georgia Institute of Technology. He is director of the
Aerospace Systems Design Laboratory where over 40 re-
search faculty and 200 graduate students conduct research
in civil aviation, propulsion and power, advanced con-
cepts, systems engineering, and defense and space.