1
How Far Two UAVs Should Be subject to
Communication Uncertainties
Quan Quan, Rao Fu, and Kai-Yuan Cai
Abstract—Unmanned aerial vehicles are now becoming in-
creasingly accessible to amateur and commercial users alike. A
safety air traffic management system is needed to help ensure
that every newest entrant into the sky does not collide with
others. Much research has been done to design various methods
to perform collision avoidance with obstacles. However, how to
decide the safety radius subject to communication uncertainties
is still suspended. Based on assumptions on communication
arXiv:2110.09391v1 [cs.RO] 18 Oct 2021
uncertainties and supposed control performance, a separation
principle of the safety radius design and controller design is
proposed. With it, the safety radius corresponding to the safety
area in the design phase (without uncertainties) and flight phase
(subject to uncertainties) are studied. Furthermore, the results
are extended to multiple obstacles. Simulations and experiments
are carried out to show the effectiveness of the proposed methods.
Index Terms—Safety radius, Swarm, Collision avoidance,
Communication, Separation principle, Unmanned aerial vehicle,
Air traffic
I. I NTRODUCTION
With the on-going miniaturization of motors, sensors and
processors, the number of Unmanned aerial vehicles(UAVs)
continues to explode and they increasingly play an integral role
in many practical applications in places where the working
environment is dangerous or human capacity is limited [1].
Hence, in increasingly busy airspace, the conflicts among
UAVs will occur frequently and become a serious problem.
UAS Traffic Management (UTM) by NASA in the USA
[2] and U-SPACE in the European Union [3] in progress
are aiming to manage UAVs for tactical self-separation and
collision avoidance [4].
Traditionally, the main role of air traffic management (ATM)
is to keep a prescribed separation between all aircraft by using
centralized control. However, it is infeasible for increasing
UAVs because the traditional control method lacks scalability.
In order to address such a problem, free flight is a devel-
oping air traffic control method that uses decentralized con-
trol [5],[6]. By Automatic Dependent Surveillance-Broadcast
(ADS-B) [7], Vehicle to Vehicle (V2V) communication [8] or
5G mobile network [9], UAVs can receive the information of
their neighboring obstacles to avoid collision. Especially in
low-altitude airspace, utilizing the existing mobile networks
will eliminate the need to deploy new infrastructure and, there-
fore, help to ensure connected UAVs are economically feasible
[13]. Communication is often considered for UAVs, especially
in the formation control, where the communication connection
establishes the network topological structure [10],[11],[12].
Q. Quan, R. Fu and K-Y. Cai are with the School of Automation Science
and Electrical Engineering, Beihang University, Beijing 100191, China (e-
mail: qq buaa@buaa.edu.cn; buaafurao@buaa.edu.cn; kycai@buaa.edu.cn).
However, it is not enough because communication protocols
are subject to message delivery delay and packet loss. A V2V
communication test on UAVs has been done by NASA for
UTM showing that the probability of packet loss will be
increased gradually to 1 after two UAVs keep about 1.3km
away [14]. A UAV typically relies on localization sensors,
monitoring systems, and/or wireless communication networks.
These sensing mechanisms will provide inaccurate position
information as a result of process delays, interferences, noise,
and quantization, jeopardizing the safety of the vehicle and
its environment [15]. If these uncertainties are not adequately
taken into consideration, then UAVs may become vulnerable to
collisions. This motivates us to study UAV collision avoidance
problem for moving obstacles [16],[17] subject to sensing
uncertainties, especially communication uncertainties.
Much research has been done to design various methods,
including path planning, conflict resolution, model predictive
control, potential field, geometric guidance, motion planning
of teams, for UAVs to perform collision avoidance with
obstacles [18],[19]. A simulation study of four typical collision
avoidance methods can be found in [20]. Most of existing
collision avoidance methods suppose that they can acquire
exact sensing information. However, in recent years, more and
more attention has been paid to collision avoidance subject
to sensing uncertainties. Two ways are followed to extend
the existing collision avoidance methods to handle sensing
uncertainties for moving obstacles.
• The major way is to predict the future obstacles’
trajectory set with uncertainty models [21], [22],
[23],[24],[25],[26]. Then, control, decision or planning
is obtained by optimization over a future time horizon to
avoid obstacles’ trajectory sets or reachable sets in the
sense of probability. This way is often for the control
methods without closed forms, like the optimization
methods mentioned above. However, as pointed in [27],
under some circumstances the optimization problem may
be infeasible due to the sensing uncertainties. Moreover,
heavy computational burden always introduces difficulties
in the design process. A discussion will be given to
show this in Section IV.C. As far as we know, fewer
communication uncertainties are considered in this way.
• The other way is to take the sensing uncertainties into
the analysis of the closed-loop of existing methods. Then,
adjust controller parameters to reject uncertainties [15].
This way is often for the control methods with closed
forms, like the potential field method. However, the
analysis will be more difficult subject to the uncertainties
like communication uncertainties. What is worse, the
 2

communication uncertainties may make the closed-loop To reply to these two questions, first, a VTOL UAV control
instability if the delay or packet loss is not compensated model and an obstacle model are proposed. The filtered
for elaborately [28]. position is defined to replace position by considering velocity.
Based on these models above, assumptions on practical uncer-
As pointed out in [29], communication (delivery delay and
tainties, like estimation error, communication delay and packet
packet loss) is a very significant and necessary factor that
loss, are considered in the broadcast information received by
should be considered in UAV swarm. The survey paper [18]
the UAV. Assumptions on control performance are assumed for
also takes communication uncertainties in collision avoidance
the design phase and the flight phase, respectively. Since only
as a challenge. To deal with sensing uncertainties including
is the distance error used, the controller can be distributed.
communication uncertainties, a complementary way of col-
Based on these assumptions, a principle of separation of
lision avoidance is proposed in this paper with a principle
control and safety radius is proposed in the design phase
that separates the safety radius design for uncertainties and
(Theorem 1). Based on this principle, the safety radius in the
controller design for collision avoidance. Here, the safety
design phase is derived. Then, the safety radius in the flight
radius design will take all sensing uncertainties into consid-
phase is further derived in the sense of estimated distance.
eration, while the controller design does not need to take
These conclusions are further extended to multiple obstacles.
uncertainties into consideration. So, the two ways mentioned
Simulations and experiments including delay and package-loss
above can also benifit from the safety radius design when
uncertainties are carried out to show the effectiveness of the
facing communication uncertainties. Intuitively, the safety
proposed method.
radius will be increased as the uncertainties are increased.
This is inspired by traffic rules in both the aviation area
[30] and the ground transportation area [31]. For example, Car (Controller inside) Obstacle
two airplanes should maintain standard en route separation
between aircraft (5 nautical miles (9.3 km) horizontal and
1,000 feet (300 m) vertical) [30]. Also, as shown in Figure
1, it is well known that two cars on highway should keep a Safety distance
certain safe distance. However, the experience to determine
the safety radius for manned and airline airplanes in high-
Fig. 1. Safe distance of cars in highway.
altitude airspace or safety distance for cars is difficult or
not at all to apply to decentralized-control UAVs in low-
As shown in Figure 1, let us recall the safe distance for cars
altitude space because of the differences in pilot manner,
in highway (1D space), where drivers in cars are controllers.
communication manner, flight manner and risk requirement.
When set the safety distance for all of us, we in fact make
On the other hand, obviously, the safety radius cannot be
an assumption on the control performance for all drivers
just the physical radius of a UAV because many uncertainties,
(controllers) by taking the drivers’ response delay and road
like estimation error, communication delay and packet loss,
condition’s uncertainties. This idea leaves uncertainties in the
should be considered. As pointed out in [18], the safety radius
safe distance. Motivated by this, we do the similar work here
design is a challenge in the presence of sensing uncertainties,
for UAVs in 3D space. In the past decades, many collision-
especially communication uncertainties.
avoidance controllers have been proposed [33],[34],[35] sim-
To this end, this paper focuses on studying the safety radius ilar to different drivers. We only need to focus on their
of decentralized-control Vertical TakeOff and Landing (VTOL)
performance rather than what they are. Compared with the
UAVs, while the deterministic collision avoidance controller
safe distance in the traditional manned traffic field, there are
could be any type as long as it satisfies certain conditions. two extras taken into consideration: i) cooperative obstacles,
VTOL ability, which enables easier grounding or holding by
which will make collision avoidance at the same time; ii) the
hovering, is an important ability that might be mandated by packet loss, namely a dynamic uncertainty. These make the
authorities in high traffic areas such as lower altitude in the
obtained conclusion not easy. The major contribution of this
urban airspace [32]. This is because VTOL drones are highly
paper is to separate the controller design and the safety radius
versatile and can perform tasks in an environment with very so that the existing methods are available again (not need to
little available airspace. The study on the safety radius is
be changed) in the presence of communication uncertainties.
divided into two phases, namely the ‘offline’ design phase
Concretely, the contributions of this paper are: i) a principle
and ‘online’ flight phase. of separation of control and safety radius proposed so that the
• In the design phase, when a UAV is on the ground, a same controller with different safety radiuses can deal with
question will arise that how far the UAV and an obstacle different communication uncertainties; ii) safety radius derived
should be kept without uncertainties in order to avoid a in both the design phase and the flight phase to reduce the
collision in the presence of uncertainties? conservatism as much as possible; iii) filtered VTOL UAV
• On the other hand, when a UAV in the sky (in the control model in the form of a single integrator which takes
presence of uncertainties), a question will arise that how the maneuverability into consideration.
far should the UAV and an obstacle be kept in the sense The remainder of this paper is organized as follows. In
of estimated distance (involve uncertainties but only can Section II, the problem is formulated, which based on a
be accessed) in order to avoid a collision? proposed VTOL UAV control model, a proposed obstacle
 3

model, and assumptions. The solutions to the problem and $PDUJLQDODYRLGDQFH

their extensions are proposed in Section III. The effectiveness GLVWDQFH
of the proposed safety radius design is demonstrated by sim-
ulation and flight experiments in Section IV. The conclusions
are given in Section V. Some details of the mathematical proof
process are given in Section VI as an appendix.
D 7ZR8$9VNHHSDZD\ZLWKRSSRVLWHYHORFLW\GLUHFWLRQ

II. P ROBLEM F ORMULATION

In this section, a VTOL UAV model and obstacle model E 7ZR8$9VJHWFORVHZLWKRSSRVLWHYHORFLW\GLUHFWLRQ

are introduced first. Then, assumptions on uncertainties and
control performance are proposed. Based on models and
assumptions, problems for the design phase objective and
flight phase objective are formulated. F 7ZR8$9VIO\ZLWKWKHVDPHGLUHFWLRQ

Fig. 2. Intuitive interpretation for filtered position

A. VTOL UAV Control Model

In a local airspace, there exists a VTOL UAV defined as Then
U = x ∈ R3 |kx − pk < rm

(1)
1
where rm > 0 is the physical radius of the UAV related to ξ̇ = ṗ + v̇
l
its physical size, and p ∈ R3 is the center of mass of the 1
UAV. Many organizations or companies have designed some = v − l (v − vc )
l
open-source semi-autonomous autopilots or offered semi- = vc . (5)
autonomous autopilots with Software Development Kits. The
semi-autonomous autopilots can be used for velocity control of The relationship among the position, the filtered position
VTOL UAVs. With such an autopilot, the velocity of a VTOL and the estimated filtered position is shown in Figure 3,
UAV can track a given velocity command in a reasonable time. where the UAV cannot access the position and the filtered
It can not only avoid the trouble of modifying the low-level position (ground truth) but only the estimated filtered position
source code of autopilots but also utilize commercial autopilots including uncertainties. From the UAV’s view, it senses itself
to complete various tasks. Based on this, the control model of and obstacle at both estimated filtered positions (dash circle)
the UAV satisfies in Figure 3. If the UAV is at a high speed or the maneuver
constant l is small, then true position p will be far from its
ṗ = v filtered position.
v̇ = −l (v − vc ) (2)
(VWLPDWHG2EVWDFOH
where l > 0, v ∈ R3 is the velocity, and vc ∈ R3 is the
(VWLPDWHG8$9
velocity command. From the model (2), lim kv (t) − vc k = 0 ro
t→∞ (VWLPDWHG
if vc is constant. The constant l, called maneuver constant ILOWHUHGSRVLWLRQ ξ̂ o
Estima rs
here, depends on the VTOL UAV and the semi-autonomous ted
Positio Filtereed (VWLPDWHG
n Erro
autopilot used, which can be obtained through flight experi- r
ILOWHUHGSRVLWLRQ ξ̂
ments. It stands for the maneuverability of the VTOL UAV. If
it is big, then v can converge to vc rapidly, vice versa. Here,
the velocity command vc (required to design) for the VTOL rs
)LOWHUHGSRVLWLRQ ξ
UAV is subject to ro on E r r
or
Positi v
)LOWHUHGSRVLWLRQ
VLWLRQ ξo Filtere
d
max kvc k ≤ vm . (3) vo tion Err
or 7
7UXHSRVLWLRQ p
P o si
7UXHSRVLWLRQ
WLRQ po 1RLVH'HOD\
3DFNHW/RVV 8$9
As shown in Figure 2, although the distances between two rm
2EVWDFOH
UAVs in the three cases are the same, namely a marginal
avoidance distance, the case in Figure 2(b) needs to carry out Fig. 3. Relation among position, filtered position, estimated filtered position,
avoidance urgently by considering the velocity. However, the where the estimated filtered position will appear in (9) and (10).
case in Figure 2(a) in fact does not need to be considered to
perform collision avoidance. With such an intuition, a filtered Remark 1. It should be noted that the model (2) is a second-
position is defined as follows: order system. By the transformation (4), the model (2) is
1 transformed into a single integrator form to simplify the further
ξ , p + v. (4) controller design and analysis. What is more, the proposed
l
 4

model (2) has taken the maneuverability into consideration. • Information from itself. Based on Assumption 1, as for
Although a commonly-used model the UAV itself, the filtered position estimate is
ṗ = vc (6) ξ̂ = ξ + ε (10)
is also a single integrator model, it does not take the maneuver-
ability into consideration. As a result, the true velocity v will because no broadcast delay and packet loss need to be
be different from vc . The difference will be big, if the UAV’s considered.
maneuverability is low. This further increases the difference • Information from obstacle. The UAV has to receive
between the true position and desired position especially in information from the obstacle via communication. As
some situations such as making an aggressive maneuver. shown in Figure 4, the UAV receives the information ξ̂o ,
which has changed by estimate noise, broadcast delay,
and packet loss.
B. Obstacle Model
In the same local airspace, there exists a moving obstacle Estimate
(it may be an aircraft or a balloon) defined as Noise
Obstacle UAV
+
O = x ∈ R3 |kx − po k < ro

+ Broadcast
Packet Loss
Delay
3 Position and
where ro > 0 is the obstacle radius, and po ∈ R is the center Velocity
of mass of the obstacle. Define
1 Fig. 4. Shared information broadcasting
ξo , po + vo
l
where vo ∈ R3 is the velocity of the obstacle. The obstacle Remark 2. The reasonability of the selected model (9) is
satisfies the following model explained as follows. First, a simple and reasonable estimation
method is adopted
max ξ̇o ≤ vo 
ξ o (t − τd ) if data packet is received
where vo > 0. This is a general model for any obstacle with ξ̂ o (t) = . (11)
ξ̂ o (t − Ts ) if data packet is lost
bounded velocity and acceleration. Let
This implies that the estimate will remain the last estimated
ξ̇ o = ao (7) value if data packet is lost. Then, according to the probability
with max kao k ≤ vo . Then (7) can be rewritten as of packet loss θ, the expected value of ξ̂ o is

ṗo = vo ξ̄ o (t) = θξ̄ o (t − Ts ) + (1 − θ) ξo (t − τd ) , t > 0 (12)

v̇o = −l (vo − ao ) . (8)  
where ξ̄ o =E ξ̂o . Roughly, the differential ξ̄ o can be written
In particular, if kvo (0)k ≤ vo and ao = vo (0) , then the
as
obstacle is moving with a constant velocity.
ξ̄ (t) − ξ̄o (t − Ts )
ξ̄˙o (t) ≈ o . (13)
Ts
C. Assumptions on Uncertainties
Assumption 1(Estimate noise). For the UAV, the position Consequently, by using (13), the algebraic transformation of
estimate during the flight is ξ + ε, where kεk ≤ b and kε̇k ≤ (12) is
vb . 1−θ 1−θ
Assumption 2(Broadcast delay & Packet loss). The obstacle ξ̄˙o (t) ≈ − ξ̄ o (t) + ξ (t − τd ) .
θTs θTs o
can be surveilled and then broadcast, or it can broadcast its
information to the UAV. The interval of receiving information Furthermore, putting uncertainties on εo , we have the differen-
for the UAV is Ts > 0, while the time delay (including the tial equation model (9). Therefore, we can replace the model
broadcast period) is 0 < τd ≤ τdm . Let θ ∈ [0, 1] be the (11) with the new model (9). In the following simulation, the
probability of packet loss, θ ≤ θm . The estimate ξ̂ o is a value reasonability of the selected model (9) will be further shown.
that the UAV gets the estimated information from the obstacle Remark 3. The broadcast delay τdm is the maximum delay
via communication with the following model we can accept, which is the worst case. By using it, the most
1−θ 1−θ conservative safety radius is derived. If the delay exceeds τdm ,
ξ̄˙o (t) = − ξ̄o (t) + ξ (t − τd ) then the packet can be considered as a loss. The time delay
θTs θTs o
ξ̂ o (t) = ξ̄ o (t) + εo , ξ̄o (0) = ξ o (−τd ) (9) and the probability of packet loss are very normal parameters
and can be easily measured for communication [14]. For most
where kεo k ≤ bo and kε̇o k ≤ vbo . UAVs, filters will be used to eliminate the high-frequency
As shown in Figure 3, the value ξ̂ (·) represents the estimated noise to preserve the low-frequency information. As a result,
filtered position. There exist two cases: b, vb , bo , vbo will be small.
 5

D. Assumption on Controller uncertainties. In the practice flight phase, eo rather than ξ̃ o

Let is only available. So, the same controller is used but with
different feedback such as
p̃o , p − po
vc = c (t, eo ) (21)
ṽo , v − vo
ξ̃o , ξ − ξ o . (14) for (2), where
eo , ξˆ − ξ̂o . (22)
For the UAV, no collision with the obstacle implies
Since the controller cannot get the truth, we only can make the
U ∩O =∅ estimated filtered position distance, namely keo k , greater than
a certain distance. Assumption 4 is only for the flight phase in
namely true position distance kp̃o k satisfies
the presence of uncertainties, describing what has happened.
kp̃o k ≥ rm + ro . (15) Assumption 4 (Control Performance in Flight Phase). There
exists a practical safety radius rs′ > 0 such that, for any
In the following, there are two assumptions about controller
obstacle with ξ̇o ≤ vm and keo (0)k ≥ rs′ + ro , a controller
design. Assumption 3 is only for the design phase, during
for (2) can make
which a controller is designed for (2) without considering
uncertainties to make the filtered position distance, namely keo (t)k ≥ rs′ + ro (23)
ξ˜o , greater than a certain distance.
where t ≥ 0.
Assumption 3 (Controller in Design Phase). Given a
designed safety radius rs > 0, with any ξ̃o (0) satisfying
E. Objective
ξ˜o (0) ≥ rs + ro , a controller
  In the ‘offline’ design phase (in numerical simulation), a
vc = c t, ξ̃o (16) question will arise that how far should the UAV and the
obstacle be kept without uncertainties offline in order to avoid
for (2) can make a collision in the presence of uncertainties?
• Design Phase Objective (Offline). Under Assumptions
ξ̃o (t) ≥ rs + ro (17)
1-3 and controller (16) for (2), the objective here is to
  determine the estimated safety radius r̂s > 0 (it is related
for any obstacle with ξ̇o ≤ vm , where c t, ξ̃o ≤ vm ,
to rs ) that
for t ≥ 0, ∀ξ̃o ∈ R3 . keo (t)k ≥ r̂s + ro (24)
Assumption 3 is equivalent to Lemma 1.
Lemma 1. If and only if holds, where t ≥ 0. Furthermore, determine the designed
safety radius rs to make (15) hold in the presence of
minxT c (t, x) ≥ (rs + ro ) vm (18) uncertainties.
x∈C
On the other hand, when the UAV in practice (in the presence
then (17) holds for any obstacle ξ̇o ≤ vm and ξ̃ o (0) ≥ of uncertainties), a question will arise that how far should the
rs + ro , where C = x ∈R3 kxk = rs + ro . UAV and the obstacle be kept in the sense of the estimated


Proof. The idea here relies on the following inequality
minxT ẋ ≥ 0. (19)
x∈C
If and only if (19) holds, then kx (t)k ≥ rs + ro for any x with
kx (0)k ≥ rs + ro . The remaining proofs are in Appendix. 
It is necessary to assume ξ̇o ≤ vm . Otherwise, in the
worst case, the collision cannot be avoided. For example, we
can choose the obstacle dynamic as
ξo − ξ
ξ̇o = ξ̇ − ǫ (20)
kξo − ξk
where ξ̇ o ≤ ξ̇ + ǫ ≤ vm + ǫ with ǫ > 0. From (20), it is
easy to see kξo (t) − ξ (t)k < rs + ro within a finite time no
matter how small ǫ is. This implies that the obstacle is chasing
after the UAV and then hits it finally.
In Assumption 3, we do not consider the uncertainties
because the communication uncertainties will be left to the de-
signed safety radius so that the same controller with different
design safety radiuses can deal with different communication
filtered position distance, namely keo k , in order to avoid a
collision?
• Flight Phase Objective (Online). Under Assumptions 1-
2,4, the second objective is to determine the practical
safety radius rs′ > 0 to satisfy (15).
Remark 4. The design phase objective is a type of the
principle of separation of control and safety radius. It can be
stated that under some assumptions, the problem of designing a
collision-avoidance controller with uncertainties can be solved
by designing a safety radius covering the uncertainties, which
feeds into a deterministic collision avoidance controller for
the system. Thus, the problem can be broken into two separate
parts, which facilitates the design. One controller with different
designed safety radiuses can deal with different communica-
tion uncertainties. In some cases, controllers may not satisfy
Assumption 3 but Assumption 4 in practice. This is because the
controller in Assumption 3 considers any initial condition, or
say the worst condition. But it is not necessary. For example,
a UAV and an obstacle are supposed to only fly along two
separate airlines without any collision avoidance control. In
 6

this case, Assumption 3 is not satisfied, but Assumption 4 may then (24) holds with r̂s = rs for any keo (0)k ≥ rs + ro ; (ii)
hold. This motivates us to consider the safety radius separately furthermore, under Assumptions 1-3, if
in the design phase and the flight phase. ˙
(rs + ro ) (vm − vb ) ≥ eTo ξ̂ o (32)
keo k=rs +ro
III. S AFETY R ADIUS D ESIGN
then (24) holds with r̂s = rs ; (iii) in particular, under
Preliminaries are given first. By using them, the estimated Assumptions 1-3, if
safety radius and the practical safety radius are designed
for the UAV and one obstacle. Furthermore, the results are vm ≥ vo + vb + vbo , (33)
extended to multiple obstacles. then (24) holds with r̂s = rs .
Proof. See Appendix. 
A. Preliminary
(VWLPDWHG2EVWDFOH
First, an important lemma is proposed. (VWLPDWHG8$9
Lemma 2. Let x (t) ∈ Rn and y (t) ∈ Rn satisfy
(VWLPDWHG
ẋ (t) = −k (t) x (t) + k (t) y (t) (25) (VWLPDWHG
ILOWHUHGSRVLWLRQ
ILOWHUHGSRVLWLRQ ˆ o
where 0 < kmin ≤ k (t) ≤ kmax . If ky (t)k ≤ ymax , kẏ (t)k ≤ ˆ
vymax , and kx (0)k ≤ ymax , then
kx (t)k ≤ ymax , t ≥ 0. (26) (a) Cooperative obstacle and UAV with opposite velocity direction

1 (VWLPDWHG2EVWDFOH (VWLPDWHG8$9
If kx (0) − y (0)k ≤ kmin vymax holds, then
kmax (VWLPDWHG
kẋ (t)k ≤ vy , t ≥ 0. (27) (VWLPDWHG
kmin max ILOWHUHGSRVLWLRQ ˆ o
ILOWHUHGSRVLWLRQ

ˆ
Proof. See Appendix. 
With Lemma 2 in hand, we have
Proposition 1. If kv (0)k ≤ vm and the model (2) is subject (b) Non-cooperative obstacle and UAV with same velocity direction
to (3), then kv (t)k ≤ vm , t ≥ 0.
Proof. It is easy from Lemma 2.  Fig. 5. A cooperative obstacle and a non-cooperative obstacle
According to Proposition 1, we have
Remark 5. Through Theorem 1, we obtain keo (t)k ≥ r̂s +
1
kṽo k ≤ rv (28) ro , based on which we can directly determine the lower bound
l of the safety radius to make UAVs safe. From now on, we do
where not need to consider the controller any more, because Theorem
vm + vo
rv = . (29) 1 has seperated the controller design and safety radius design.
l
So, we call Theorem 1 as separation theorem.
In the following, a relationship between the true position error Remark 6. Let us consider a cooperative obstacle in the best
p̃o and the filtered position error ξ̃ o is shown. Proposition 2 case shown in Figure 5(a) and a non-cooperative obstacle in
implies that the UAV and the obstacle will be separated largely the worst case shown in Figure 5(b). The cooperative obstacle
enough if their filtered position distance is separated largely can be considered as another UAV with the same controller,
enough. namely
Proposition 2. For the VTOL UAV and the obstacle, if and ao = c (t, −eo )
only if the filtered position error satisfies
p where, for simplicity, −eo is used for another UAV’s feedback
ξ̃o (t) ≥ r2 + rv2 (30) approximately. The obstacle (another UAV) will, in turn, take
the UAV as its “obstacle” and will make collision avoidance
and kp̃o (0)k ≥ r, then kp̃o (t)k ≥ r, where t > 0. The simultaneously as well. According to Lemma 1, the following
v T vo
relationship “=” holds if kvkkv ok
= −1. inequality
Proof. See Appendix. 
−eTo c (t, −eo ) keo k=rs +ro
≥ (rs + ro ) vm
B. Separation Principle still holds. Since ξ̇ o = c (t, −eo ) , we have
For the design phase, the principle of separation of con-
troller is stated in Theorem 1. eTo ξ̇o ≤ − (rs + ro ) vm < 0.
keo k=rs +ro
Theorem 1 (Separation Theorem). Suppose that the UAV
is with model (2) under Assumptions 1-2. Then (i) if and only ˙ ˙
If ξ̂ o ≈ ξ̇o , then eTo ξ̂ o < 0. Therefore, (32) is
if keo k=rs +ro

˙
 satisfied in most cases without the requirement (33). Figure
eTo ξ̇ − eTo ξ̂ o ≥ (rs + ro ) vb (31) 5(a) shows the best case that the UAV and obstacle can keep
keo k=rs +ro
 7

away simultaneously with the opposite direction. Intuitively, If (37) holds, then
no limitation will be put on vo . Figure 5(b) shows the worst 2 2
(rs + ro − re ) ≥ (rm + ro ) + rv2
case that the UAV and obstacle move simultaneously in the
same direction. In this case, the requirement (33) implies that namely q
the UAV should have a faster speed than the obstacle’s speed. 2
ξ̃ o (t) ≥ (rm + ro ) + rv2 .

C. Safety Radius Design According to Proposition 2, we have

Based on rm , ro > 0, we will further determine rs > 0 in kp̃o (t)k ≥ rm + ro .
order to avoid a collision in the presence of uncertainties.
For this purpose, we need to analyze the relationship between 
the filtered position distance and the true position distance With Theorem 2 in hand, the solution to the flight phase
(Proposition 2 has done), and the relationship between the objective is easy to get.
filtered position distance and the estimated filtered position Theorem 3. Under Assumptions 1-2,4, if the practical safety
distance (Proposition 3 will show), as shown in Figure 6. radius satisfies
q
2
rs′ ≥ (rm + ro ) + rv2 + re − ro
then (15) holds.
> rm + ro
True Position Distance00000000
Proof. It is similar to the proof of Theorem 2. 

Relationship between Filtered

D. Extension to Multiple Obstacles
Position Distance and True Position The results above can be extended to multiple obstacles as
Distance (Proposition 2)
well. There are M obstacles
Oo,k = x ∈ R3 |kx − po,k k ≤ ro


Relationship between Filtered Position

Distance and Estimated Filtered Position where po,k ∈ R3 is the center position of the kth obstacle,
Distance (Proposition 3) vo,k = ṗo,k ∈ R3 is the velocity of the kth obstacle, k =
1, · · · , M . Define
rs ?
1
Estimated Filtered Position Distance ξo,k , po,k + vo,k
> rs + ro
l
ξ̃o,k , ξ − ξo,k .

These obstacles satisfy max ξ̇ o,k ≤ vo , k = 1, · · · , M . For

Fig. 6. Relationship from true position distance to estimated filtered position
distance the UAV, no collision with the multiple obstacles implies

Proposition 3. Under Assumptions 1-2, given any r > 0, if U ∩ Oo,k = ∅ (38)

keo (t)k ≥ r + re , t ≥ 0 (34) where k = 1, · · · , M. To extend the conclusions in Theorems

1-2 to multiple moving obstacles, we have Assumptions 2’,3’
then to replace with Assumptions 2,3 in the following.
ξ̃o (t) ≥ r, t ≥ 0 (35) Assumption 2’ (Broadcast delay & Packet loss). The kth
obstacle can be surveilled and then broadcast, or it can
where
θm T s broadcast its information to the UAV. The interval of re-
re = vo + vo τdm + b + bo (36) ceiving information for the UAV is Ts > 0, while the time
1 − θm
delay (including the broadcast period) of the kth obstacle is
Proof. See Appendix.  0 < τd,k ≤ τdm . Let θk ∈ [0, 1] be the probability of packet
With Lemma 2, we start to determine the lower bound of loss for the kth obstacle, θk ≤ θm < 1 . The estimate ξ̂ o,k is
the safety radius rs in the design phase. a value that the kth UAV gets the estimated information from
Theorem 2. Under Assumptions 1-3, if the designed safety the obstacle via communication with the following model
radius satisfies
1 − θk 1 − θk
ξ̄˙o,k (t) = −
q
2 ξ̄ o,k (t) + ξ (t − τd,k )
rs ≥ (rm + ro ) + rv2 + re − ro , (37) θk T s θk Ts o,k
then condition (15) holds. ξ̂o,k (t) = ξ̄ o,k (t) + εo,k , ξ̄o,k (0) = ξo,k (−τd,k ) (39)
Proof. According to Proposition 3, if where kεo,k k ≤ bo and kε̇o,k k ≤ vbo , k = 1, · · · , M.
keo (t)k ≥ rs + ro The avoidance case with multiple obstacles is complex.
Under some initial conditions, the UAV cannot avoid collision
then with obstacles no matter what a controller uses, such as a case
ξ̃ o (t) ≥ rs + ro − re . shown in Figure 7. For such a purpose, we define a set F for
 8

2EVWDFOH 2EVWDFOH In view of (9), according to Lemma 1, we have

ξ̄˙o,k ≤ ξ̇ o,k (t − τd,k )

≤ vo .
Then, (44) is bounded as
′
ξ̇ o,k ≤ ξ̄˙o,k + kε̇o,k k + kε̇k
≤ vo + vbo + vb
2EVWDFOH 2EVWDFOH

where Assumptions 1,2’ are utilized. Therefore, if (33) holds,

Fig. 7. UAV surrounded by four obstacles then (43) holds with r̂s = rs . 
Remark 7. The introduction to the set F is to make the
problem completed, which is out of the scope of this paper.
multiple obstacles’ and UAV’s initial conditions in Assumption
How to find the set F is an interesting problem, which can
3’.
Assumption 3’. Given a designed safety radius rs > 0, with
be
 formulated as: given  a T > 0, the initial condition set
  ξ̃o,1 (0) , · · · , ξ̃ o,M (0) ∈ F is a set that can make
any ξ̃o,1 (0) , · · · , ξ̃ o,M (0) ∈ F , a controller
 
  max min ξ̃o,1 (t) , · · · , ξ̃ o,M (t) ≥ rs + ro , 0 ≤ t ≤ T
vc = c t, ξ̃o,1 , · · · ,ξ̃o,M (40) vc ao,1 ,···ao,M
˙
s.t. ξ̃ o,k = vc − ao,k , kvc k ≤ vm , kao,k k ≤ vo , k = 1, · · · , M.
for (2) can make
Interested readers can take the problem as the feasibility of the
ξ̃ o,k (t) ≥ rs + ro (41) pursuit-evasion game problem with multiple entities chasing a
single target or prey [36] or group chase and escape problem
for obstacles with ξ̇ o,k ≤ vo , where [37].
  To extend the conclusions in Theorem 3 to multiple moving
c t, ξ̃ o,1 , · · · ,ξ̃ o,M ≤ vm , for t ≥ 0, k = 1, · · · , M. obstacles, we have Assumption 4’ to replace with Assumption
Theorem 4. Suppose that the UAV is with model (2) under 4 in the following.
Assumptions 1-2. Then (i) if and only if Assumption 4’. There exists a practical safety radius rs′ >

˙
 0 such that, for obstacles with ξ̇o,k ≤ vo and keo,k (0)k ≥
eTo,k ξ̇ − eTo,k ξ̂ o,k ≥ (rs + ro ) vb (42)
keo,k k=rs +ro rs′ + ro , a controller for (2) can make
then keo,k (0)k ≥ rs′ + ro
keo,k (t)k ≥ rs + ro (43)
where k = 1, · · · , M .
for any (eo,1 (0) , · · · , eo,M (0)) ∈ F where eo,k , ξ̂ − ξ̂ o,k , In the flight phase, the results in Theorem 3 are extended
k = 1, · · · , M. (ii) In particular, under Assumptions 1,2’,3’ , in the following Theorem 5 for multiple obstacles.
if (33) holds, then (43), where k = 1, · · · , M. Furthermore, if Theorem 5. Under Assumptions 1,2’,4’, if the practical
rs satisfies (37), then (38) holds for k = 1, · · · , M . safety radius satisfies
Proof. Proof of Conclusion (i) does not rely on Assumption q
2
3’, which is similar to Conclusion (i) of Theorem 1. So, we rs′ ≥ (rm + ro ) + rv2 + re − ro
omit it. Let us prove Conclusion (ii). The controller (40) is
then (38) holds for k = 1, · · · , M .
rewritten as
Proof. It is similar to the proof of Theorem 3. 
vc = c (t, eo,1 , · · · , eo,M )
= c t, ξ − ξ′o,1 , · · · , ξ − ξ′o,M

IV. S IMULATION AND E XPERIMENTS
A. Simulation
where ξ ′o,k , ξ̂ o,k − ε. New obstacle Ok′ with filtered position
In the first two simulations for non-cooperative obstacles,
ξ ′o,k are taken into consideration, where k = 1, · · · , M .. If
′ with the proposed separation principle (Theorems 1,4) and
ξ (0) − ξ′o,1 (0) , · · · , ξ (0) − ξ ′o,M (0) ∈ F and ξ̇ o,k ≤


the designed safety radius (Theorem 2), we will show that
vo , then the condition (33) is necessary in order to make avoidance
ξ − ξ ′o,k ≥ rs + ro subject to uncertainties. However, for cooperative obstacles
which can make avoidance simultaneously, the condition (33)
namely (43) holds, according to Assumption 3’, k = is not necessary, which is shown in the last simulation. The
′
1, · · · , M . The left problem is to study the condition ξ̇o,k ≤ results of Theorems 3,5 can be observed directly from these
vm . The derivative ξ ′o,k is results of the following simulations by choosing r̂s = rs .
A video about simulations and experiments is available on
′
ξ̇ o,k = ξ̄˙o,k + ε̇o,k − ε̇. (44) https://youtu.be/MawyB3eoZQ0 or http://t.cn/A6ZD7otD.
 9

Case * b(m) bo (m) vb (m/s) vbo (m/s) τd (s) θ rs (m)

Case A 0 0 0 0 0 0 5.30
be noted that keo (t)k ≥ rs + ro holds consistent with
Case B 3 1 3 1 1 10% 14.30 conclusion (iii) in Theorem 1 during the flight, although
Case C 5 2 6 5 2 20% 22.31 ξ̃ o (t) < rs + ro after t = 5s because of uncertainties.
TABLE I
D IFFERENT COMMUNICATION PARAMETERS Since Case C does not satisfy the condition of (33), as
shown in Figure 10(c), keo (t)k < rs + ro at time about
11.25s.
x-axis direction error

rs

 rm
ro
 (VWLPDWHG3RVLWLRQE\8$9


-3 -2 -1 0 1 2 3

y-axis direction error

        
t = s






-3 -2 -1 0 1 2 3 


Fig. 8. Statistical property of obstacle estimate noise



        
t = 5s
1) Simulation with One Non-Cooperative Obstacle: 

• Simulation Setting. As shown in Figure 9, a scenario that 

one static UAV makes avoidance with one moving non- 

cooperative obstacle is considered. The simulation param- 
eters are set as follows. The UAV with a physical radius 
T
rm = 5m is at p (0) = [0 0 100] m initially. The UAV’s

maneuver constant is l = 5, and the minimum speed

vm = 10m/s. The obstacle is at po (0) = [40 0 100]T m         
t = 10s
initially with radius ro = 10m and a constant velocity 
T
vo = [−5 0 0] m/s. The interval of receiving information 
for the UAV is Ts = 0.01s. Communication uncertainty 
parameters are set as Table I, where only Case A has no

uncertainties. We make Case B satisfy the condition of

(33) but Case C not intentionally. The designed safety

radiuses are all chosen according to (37) in Theorem 2.
• Assumption Verification. The comparison of model (9) 
        
in Assumption 2 with the model (11) is studied by taking t = 15s

Case B as an example, where the value ξ̂ o is from (11)

taking as the ground truth and ξ̄o (t) from (9). The two
models with the same communication parameters have
the same input ξ o . Let us study the noise εo = ξ̂ o −
ξ̄ o . As shown in Figure 8, for the UAV, the noise εo
is bounded, moreover, obeying the normal distribution
by Kolmogorov-Smirnov test. Therefore, Assumption 2
is reasonable.
• Safety Radius Verification. Under the initial conditions
above and an obstacle avoidance controller, the true posi-
tion distance kp̃o (t)k and the estimated filtered position
distance keo (t)k between the UAV and the obstacle are
shown in Figure 10. Figure 10(a) corresponding to Case A
shows that Assumption 3 is satisfied. The results observed
from Figure 9 and Figure 10(b) corresponding to Case B,
are consistent with conclusion (iii) of Theorem 1 and the
result in Theorem 2. As shown in Figure 9, it should
Fig. 9. Positions of UAV and obstacle at different times in Case B of
numerical simulation
2) Simulation with Multiple Non-Cooperative Obstacles:
• Simulation Setting. As shown in Figure 11, a sce-
nario that one UAV makes avoidance with three moving
non-cooperative obstacles is considered. These obsta-
cles can avoid each other expect for the UAV. The
simulation parameters are set as follows. The initial
T
position of the UAV is set as p (0) = [0 40 100] m
with radius rm = 5m; the initial positions of obstacles
are set as po,1 (0) = [−40 − 40 100]T m, po,2 (0) =
T T
[0 − 40 100] m, po,3 (0) = [40 − 40 100] m with ra-
dius ro = 10m and the velocity vo,i = i + 2m/s,
i = 1, 2, 3. The others about the UAV and uncertainties
are the same as those in the last simulation.
 10

7UXH3RVLWLRQ'LVWDQFH P (VWLPDWHG)LOWHUHG3RVLWLRQ'LVWDQFH P 60 60

 %HWZHHQ8$9DQGREVWDFOH 
40 40
  


  20 20
 rs

       rm
 rN rP  0 ro 0
rs ro
  Estimated Position by UAV
           
D &DVH$
-20 -20

 
 -40 -40
   
 -60
  -60 -40 -20 0 20 40 60 -60-60 -40 -20 0 20 40 60
    
  W 0V W V
 rN rP rs ro
60 60
 
           
E &DVH% 40 40

  3
32 20 20
  32 32.
  11 1 11 0 0
 
 
rs ro
 rN rP -20 -20
 
           
F &DVH& -40 -40

-60 -60
Fig. 10. The position distance and estimated filtered position distance between -60 -40 -20 0 20 40 60 -60 -40 -20 0 20 40 60
a UAV and one obstacle in numerical simulation W 0V W V

Fig. 11. Positions of UAV and three obstacles at different time in Case B of
• Safety Radius Verification. Under the initial conditions numerical simulation
above and an obstacle avoidance controller, the mini-
7UXH3RVLWLRQ'LVWDQFH P (VWLPDWHG)LOWHUHG3RVLWLRQ'LVWDQFH P
mum true position distance min kp̃o,i (t)k and the  
i∈{1,2,3} %HWZHHQ8$9DQGREVWDFOH
 

estimated filtered position distance min keo,i (t)k be-  
i∈{1,2,3}   
   
tween the UAV and the obstacle are shown in Figure 12. 
 rN rP  rs
ro
 
Figure 12(a) corresponding to Case A shows that Assump-                
D &DVH$
tion 3’ is satisfied. The results observed from Figure 11  

and Figure 12(b) corresponding to Case B, are consistent  
 
with conclusion (ii) of Theorem 4. As shown in Figure      
  

11, it should be noted that min keo (t)k ≥ rs + ro  rN rP
rs ro
i∈{1,2,3}  
               
holds consistent with conclusion (ii) in Theorem 4 during E &DVH%
  
the flight, although min ξ̃ o,i (t) < rs + ro about 
i∈{1,2,3}   
    
t = 10s because of uncertainties. Since Case C does not   
  rs ro
satisfy the condition of (33), as shown in Figure 12(c),  rN rP
                
min keo,i (t)k < rs + ro at time about 6s. F &DVH&
i∈{1,2,3}
3) Simulation with Multiple Cooperative Obstacles: Fig. 12. Minimum position distance and estimated filtered position distance
• Simulation Setting. As shown in Figure 13, a scenario from UAV to three non-cooperative obstacles in numerical simulation
that one UAV makes avoidance with three moving co-
operative obstacles is considered. The UAV and these above and an obstacle avoidance controller, the minimum
obstacles can avoid each other. The simulation parameters true position distance min kp̃o,i (t)k and the esti-
are set as follows. The initial position of the UAV is set i∈{1,2,3}
as p (0) = [−40 40 100]T m with radius rm = 5m. The mated filtered position distance min keo,i (t)k from
i∈{1,2,3}
UAV’s maneuver constant is l = 5, and the maximum the UAV to these obstacles are shown in Figure 14. Since
speed vm = 5m/s. The interval of receiving information the obstacles and the UAV can make collision avoidance
for the UAV is Ts = 0.01s. The initial positions of with each other, the separation principle still holds even if
T
obstacles are set as po,1 (0) = [40 40 100] m, po,2 (0) = vm = vo,3 . For a simple case, Remark 5 has explained the
[40 − 40 100] m, po,3 (0) = [−40 − 40 100]T m with
T
reason. Consequently, min keo (t)k ≥ rs + ro holds
radius ro = 10m and their velocities vo,i = i + 2m/s, i∈{1,2,3}
during the flight observed from Figure 13.
i = 1, 2, 3. Communication uncertainty parameters are
set as Case B in Table I. The designed safety radius is
chosen as rs = 14.14m according to (37) in Theorem 2. It B. Experiments
is worth noting that the condition of (33) does not satisfy A motion capture system called OptiTrack is installed, from
because of vm = vo,3 . which we can get the ground truth of the position, velocity
• Safety Radius Verification. Under the initial conditions and, orientation of each multicopter. The laptop is connected
 11

60 60 
%HWZHHQ8$9DQGREVWDFOH

7UXH3RVLWLRQ'LVWDQFH P
40 40 

20 20 
8$9 2EVWDFOH
rs

0
rm
0 
ro
(VWLPDWHG3RVLWLRQE\8$9 
2EVWDFOH 2EVWDFOH -20 
-20 rm ro


-40 -40        
W VHF
60 60

(VWLPDWHG)LOWHUHG3RVLWLRQ'LVWDQFH P
-60 -40 -20 0 20 40 6 -60 -40 -20 0 20 40 6 
W 0V W 0V %HWZHHQ8$9DQGREVWDFOH
60 60


40 40 

20 20 

 rs ro
0 0 67

       
-20 -20
W VHF

-40 -40

60 60 Fig. 15. The position distance and estimated filtered position distance between
-60 -40 -20 0 20 40 6 0 -40 -20 0 20 40 6 a UAV and one hovering obstacle in the flight experiment
W 0V W V

Fig. 13. Positions of UAV and three obstacles at different times in Case B
of numerical simulation
UAV makes avoidance with one non-cooperative hovering
obstacle is considered. The experiment parameters are set
 %HWZHHQ8$9DQGREVWDFOH
as follows. The UAV with physical radius rm = 0.2m is
7UXH3RVLWLRQ'LVWDQFH P

at p (0) = [1.5 0 1]T m initially. The UAV’s maneuver


constant is l = 2, and the maximum speed vm = 0.1m/s.
T
 The obstacle is at po (0) = [−0.2 0 1] m initially with
 rm ro
 radius ro = 0.2m. The interval of receiving information
 for the UAV is Ts = 0.01s. Communication uncertainty

       parameters are set as b = 0.10m, bo = 0.03m, vb =
W VHF
0.08m/s, vbo = 0.01m/s, τd = 1s, θ = 10%. The condition
(VWLPDWHG)LOWHUHG3RVLWLRQ'LVWDQFH P

of (33) in Theorem 2 is satisfied in this scenario. The


 %HWZHHQ8$9DQGREVWDFOH

safety radius rs = 0.47m is designed with such defined
  parameters according to Theorem 2.


• Safety Radius Verification. Under the initial condi-
      


rs ro

 distance from the UAV to the obstacle are shown in
 
Fig. 14. Minimum position distance and filtered position distance from UAV
to three cooperative obstacles in numerical simulation
to these multicopters and OptiTrack by a local network,
providing the proposed controller and a real-time position
plotting module. In the first experiment for a non-cooperative
obstacle, with the proposed separation principle (Theorem 1)
and designed safety radius (Theorem 2), we will show that the
condition (33) can make avoidance subject to uncertainties.
However, for cooperative obstacles which can make avoidance
simultaneously, the condition (33) is not necessary, which is
shown in the last two experiments. The results of Theorems 3,5
can be observed directly from these results of the following
experiments by choosing r̂s = rs .
1) Experiment with One Non-Cooperative Hovering Obsta-
cle:
• Experiment Setting. An experiment scenario that one
tions above and an obstacle avoidance controller, the
 true position distance and the estimated filtered position
    
W VHF Figure 15. The positions of multicopters during the whole
flight experiment are shown in Figure 16. The UAV can
complete its route at about 77s, keeping a safe distance
from the obstacle without conflict. This is consistent with
the separation principle (Theorem 1) with the designed
safety radius (Theorem 2).
2) Experiment with One Cooperative Moving Obstacle:
• Experiment Setting. An experiment scenario that one
static UAV makes avoidance with one moving cooperative
obstacle is considered. The experiment parameters are set
as follows. The UAV with physical radius rm = 0.2m is at
p (0) = [1.5 0 1]T m initially. The UAV’s maneuver con-
stant is l = 2, and the maximum speed vm = 0.1m/s. The
T
obstacle is at po (0) = [−1.5 0 1] m initially with radius
ro = 0.2m and vo = 0.1m/s. The interval of receiving
information for the UAV is Ts = 0.01s. Communication
uncertainty parameters are set as b = 0.2m, bo = 0.1m,
vb = 0.08m/s, vbo = 0.01m/s, τd = 2s, θ = 30%. The
safety radius rs = 0.71m is designed with such defined
 12

1
ObstacůĞ UAs
0.5

-0.5 5
rs
%HWZHHQ8$9DQGREVWDFOH

7UXH3RVLWLRQ'LVWDQFH P
-1 rm
ro 4
-1.5
Estimated Position by UAV

-2 3
W V -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
1
2
ObstacůĞ 0.5 1.1305

0 1 rm ro
UAs
0
-0.5
0
-1 0 5 10 15 20 25 30 35 40 45
W VHF
-1.5

(VWLPDWHG)LOWHUHG3RVLWLRQ'LVWDQFH P
-2 5
W V -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
1 %HWZHHQ8$9DQGREVWDFOH
ObstacůĞ 4
0.5

UAs 0
3
-0.5
2
-1 1.1138
rs ro
-1.5 1

-2
W V -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
1 0
0 5 10 15 20 25 30 35 40 45
UAs ObstacůĞ 0.5
W VHF
0

-0.5
Fig. 17. The position distance and estimated filtered position distance between
-1
a UAV and one moving obstacle in the flight experiment
-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
W V

Fig. 16. Positions of a UAV and a non-cooperative hovering obstacle at

different times in the flight experiment


parameters by Theorem 4. UsϮ 
rs

rm
hsϭ
ro
• Safety Radius Verification. Under the initial condi- 
Estimated Position by UAV

tions above and an obstacle avoidance controller, the 

true position distance and the estimated filtered position 


distance from the UAV to the obstacle are shown in

Figure 17. The positions of multicopters during the whole W V       

flight experiment are shown in Figure 18. Since the UsϮ

obstacle and the UAV can make collision avoidance with Usϭ

each other, the separation principle still holds even if 

vm = vo . Remark 5 has explained the reason. This is 

consistent with the separation principle (Theorem 1) with 


the designed safety radius (Theorem 2). W V       

3) Experiment with Multiple Cooperative Obstacles: UsϮ 
Usϭ
• Experiment Setting. An experiment scenario that one 

UAV makes avoidance with three moving cooperative 


obstacles is considered. The experiment parameters are

set as follows. The UAV with a physical radius rm
T 
= 0.2m is at p (0) = [−1 1 1] m initially. The UAV’s W V       

maneuver constant is l = 2, and the maximum speed 

vm = 0.1m/s. The obstacles are at po,1 (0) = [1 1 1]T m,

Usϭ UsϮ

T T
po,2 (0) = [1 − 1 1] m, po,3 (0) = [−1 − 1 1] m 


initially with radius ro = 0.23m and vo = 0.1m/s.

The interval of receiving information for the UAV is

Ts = 0.01s. Communication uncertainty parameters are W V       

set as b = 0.012m, bo = 0.01m, vb = 0.012m/s,

vbo = 0.01m/s, τd = 0.1s, θ = 1%. The safety radius Fig. 18. Positions of UAV and a moving cooperative obstacle at different
times in the flight experiment
rs = 0.23m is designed according to Theorem 2.
• Safety Radius Verification. Under the initial conditions
above and an obstacle-avoidance controller, the minimum
 13

2


50m 0
80
 1.5 2 1 0
60 80
 1 60
50m 50m 40
0.5 40
rs
20 20
0 rm 50m 0
ro
3 4 0     0 00     0
-0.5 time delay = 0s time delay = s
-1 D connected digraphDQGGHVLUHG E GLVWDQFHEHWZHHQ8$9V
Iormation of UAVs
-1.5
-2
-2 -1 0 1 2
W V
2
Fig. 20. Performance of formation control method subject to communication


uncertainties
1.5
 1
 0.5 Exit Exit
0
-0.5 UAV2' UAV1'
-1 Boundary Boundary
-1.5 UAV2'
UAV1 UAV2 UAV2
-2
W V -2 -1 0 1 2
2
1.5
 UAV1
 1

0.5 UAV1'
0

-0.5
-1
-1.5
(a) UAV with a small safety radius (b) UAV with a large safety radius
-2
W V -2 -1 0 1 2
2
Fig. 21. Positions of two UAVs without and with the designed safety radius
1.5
1

  0.5
0
-0.5
-1
  -1.5
-2
W V -2 -1
Fig. 19. Positions of four UAVs at different time in flight experiment
true position distance min kp̃o,i (t)k and the esti-
i∈{1,2,3}
mated filtered position distance min keo,i (t)k from
i∈{1,2,3}
the UAV to these obstacles are shown in Figure 19. Since
these obstacles and the UAV can make collision avoid-
ance with each other, the separation principle still holds
even if vm = vo . Consequently, min keo (t)k ≥ rs +ro
i∈{1,2,3}
holds during the flight observed from Figure 19.
C. Discussion
In Assumption 3, we proposed some requirements on the
controller performance without considering the communica-
tion uncertainties. For multi-agent systems, the current con-
trol methods mainly include the optimal trajectory method
and forcefield-based method. By simulation and analysis, the
performance of different control methods subject to commu-
nication uncertainties is shown.
1) Forcefield-Based Control without Safety Radius Design:
The impact of communication uncertainties on forcefield-
based control is first shown by an example of formation
control. The objective of the formation control method re-
quires agents maintaining a certain formation. In [38], ex-
isting results of the formation control were introduced and
categorized. We choose a simple and classical displacement-
based control method to show the ability of the formation
control method subject to communication uncertainties. The
UAV’s maneuver constant is l = 5, and the maximum
speed vm = 10m/s. The initial positions of UAVs are set
as p1 (0) = [−40 40 100]T m, p2 (0) = [40 40 100]T m,
T T
0 1 2 p3 (0) = [40 − 40 100] m, p4 (0) = [−40 − 40 100] m.
The communication topology and desired formation of UAVs
are shown in Figure 20(a). For simplicity, we only consider
the impact of time delay on formation. As shown in Figure
20(b), the UAVs cannot converge to the desired formation in
a short time subject to time delay. The result is similar to
[39], which indicates that this objective is difficult to achieve
if the communication uncertainties are not compensated for
elaborately in formation controllers.
Estimating-and-then-compensating is a way to deal with
uncertainty. But if the noise, delay, or packet loss is not
compensated for elaborately, phenomena like deadlock will
happen. Let us consider a simple but particular example that
two UAVs pass a trapezoid tunnel. As shown in Figure 21(a),
in the presence of uncertainties, UAV1 considers UAV2 at
the position of UAV2’, while UAV2 considers UAV1 at the
position of UAV1’. In this case, a deadlock will exist, namely,
each one cannot pass the exit. Even if a deadlock does not
exist, these uncertainties will slow down the movement of the
swarm. The proposed safety radius for uncertainties can solve
this problem by separating the two UAVs large enough as
shown in Figure 21(b).
2) Calculation Speed Analysis of Optimal Trajectory
Method & Forcefield-Based Method with Safety Radius De-
sign: We discuss the performance of different control methods
with the safety radius design. The objective of the opti-
mal trajectory method requires optimal solutions of length,
time, or energy of path, which leads to longer calcula-
tion time of the online path-planning problem. In the sim-
ulation, we compare the online path-planning calculation
 14

speed of an optimization-based algorithm with the forcefield- Based on models and assumptions, problems are formulated
based method by MATLAB. In [40], a path-planning al- to determine the designed safety radius in the design phase
gorithm using Bezier curves with the open-source code objective and the practical safety radius in the flight phase
at https://github.com/byuflowlab/uav-path-optimization is pro- objective. For the first objective, the principle of separation
posed, which can find the optimal solutions to the offline of control and safety radius (Theorem 1) is proposed. With
and online path-planning problem. We design a scenario that this principle, the designed safety radius is determined in
contains 10 UAVs at the same altitude with rs = 5m. The Theorems 2,4. Then, the practical safety radius is determined
initial position of 1st UAV p1 (0) = [0 0 100]T m, while the in Theorems 3,5. By the proposed methods, a UAV can
other UAVs are distributed randomly in a 100m × 100m space keep a safe distance from other obstacles during the whole
with a constant velocity [1 0 0]T . For two different algorithms, flight. This is very necessary to guarantee flight safety in
we design 10 sets of random initial positions for the other practice. Simulations and experiments are given to show the
UAVs, run the simulation on the same computer, and record effectiveness of the proposed method from the functional
the average calculation time when the 1st UAV arrives at its requirement and the safety requirement.
T
destination [100 0 100] m. Figure 22 shows the calculation
speed performance with respect to the density by changing VI. A PPENDIX
the safety radius and the number of UAVs separately. As
A. Proof of Lemma 1
shown in Figure 22, for the same airspace, if the number
of UAVs increases or the safety radius of UAVs gets larger, (i) Proof of Sufficiency. Since
the calculation speed of the optimization-based algorithm will
ξ̃˙o = c t, ξ̃o − ξ̇ o
 
be decreased rapidly because the probability of constraint
being triggered is increasing, which brings more complex we have
ξ̃o ξ̃˙o = ξ̃o c t, ξ̃ o − ξ̇ o .
T T
   
calculations. On the contrary, the forcefield-based method (45)
can better deal with such an online path-planning problem.
However, the optimal trajectory method is better to deal with If (17) is violated, due to the continuity of ξ̃ o , there must exist
the offline path-planning problem. T 2
a time t = t1 such that ξ̃ o (t1 ) ξ̃ o (t1 ) = (rs + ro ) . Since
100 ξ̇ o ≤ vm , (45) becomes
)RUFHILHOGEDVHGPHWKRG 89.965
2SWLPDOWUDMHFWRU\PHWKRG
ξ̃o ξ̃˙o
T T
 
80
≥ ξ̃ o c t1 , ξ̃ o − ξ˜o (t1 ) ξ̇ o (t1 )
calculation time (s)

66.092 t=t1
60 T
 
≥ ξ̃ o c t1 , ξ̃ o − (rs + ro ) vm .
40
32.684
If (18) holds, then ξ̃ o ξ̃˙o
T
22.849 ≥ 0. This implies that
20 t=t1
T
1.302 1.311 1.319 1.334 ξ̃ o (t1 ) ξ̃ o (t1 ) will not be decreased any more. So, (17)
0
2 3 4 5
cannot be violated for any ξ̃ o (0) ≥ rs + ro . (ii) Proof of
6afety radius of UAVs (m)
100 Necessity. This necessary condition is proved by contradiction.
)RUFHILHOGEDVHGPHWKRG
89.965 Suppose that there exists an x∗ ∈ C such that
2SWLPDOWUDMHFWRU\PHWKRG
80
x∗T c (t, x∗ ) = (rs + ro ) vm − ǫ
calculation time (s)

60 < (rs + ro ) vm
47.228
40
where ǫ > 0. We will show that (17) will not hold for any
obstacle ξ̇ o ≤ vm and ξ˜o (0) ≥ rs + ro . Let ξ̃o (0) =
22.218
20 x∗ and
vm
6.386 ξ̇ o (0) = ξ̃ (0) .
0
0.954 1.033 1.175 1.334 rs + ro o
10 20 30 40
vm
1umber of UAVs So, ξ̇ o (0) = rs +ro ξ̃o (0) = vm . In this case, (45)
becomes
Fig. 22. Calculation speed of different algorithms vm
ξ̃ ξ̃˙
T T
o o = (rs + ro ) vm − ǫ − ξ̃ (0) ξ̃ o (0)
t=0 rs + ro o
V. C ONCLUSIONS = −ǫ < 0.
How to decide the safety radius taking communication This implies that ξ̃o (t) will be further decreased around
uncertainties into consideration is studied in this paper. First, a
VTOL UAV model and obstacle model are introduced. Then, t = 0, namely there exists a t = t2 such that ξ̃o (t2 ) <
some assumptions of communication and control are made, rs + ro . This contradicts with (17). So, (18) is also necessary.
including estimated noise, broadcast delay, and packet loss. 
 15

B. Proof of Lemma 2 C. Proof of Proposition 2

T
First, multiplying x (t) on the left side of (25) results in (i) Proof of sufficiency. Let

xT (t) ẋ (t) = −k (t) xT (t) x (t) + k (t) xT (t) y (t) . (46) p = p̃To p̃o
1
δ = ξ̃ To ξ˜o − 2 ṽTo ṽo .
Since l
1 dxT (t) x (t) According to (14), we have
xT (t) ẋ (t) =
2 dt 
1
T 
1

T
1 d kx (t)k2 d kx (t)k ξ̃o ξ̃ o = p̃o + ṽo
l
p̃o + ṽo
l
= = kx (t)k (47)
2 dt dt 1 T 2 T
T
= p̃o p̃o + 2 ṽo ṽo + ṽo p̃o . (50)
the equation (46) becomes l l
Since
d kx (t)k 1
= −k (t) kx (t)k + k (t) xT (t) y (t) . ṗ = 2p̃To ṽo
dt kx (t)k
using the equation (50), we further have
If ky (t)k ≤ ymax , then
ṗ = −lp + lδ. (51)
d kx (t)k
≤ −k (t) kx (t)k + k (t) ymax .
dt The solution p (t) can be expressed as
Let z , kxk − ymax . Then
Z t
p (t) = e−lt p (0) + e−l(t−s) lδ (s) ds. (52)
0
dz (t)
≤ −k (t) z (t) . With (28) in hand, if condition (30) is satisfied, then
dt
We consider another equation that T 1 T
δ (t) = ξ̃o ξ̃ o − ṽ ṽo ≥ r2 .
l2 o
dz ′ (t)
= −k (t) z ′ (t) , z ′ (0) = z (0) . Since kp̃o (0)k > r, we have p (0) > r2 . The solution in (52)
dt
satisfies
The solution to the equation above is Z t
−lt 2 2
Z t p (t) ≥ e r + r e−l(t−s) lds
0
−k(s)ds
′ = r2 .
z (t) = e 0 z (0) . (48)
T
Based on it, we have kp̃o (t)k ≥ r, where t ≥ 0. If kvkkv
v vo
=
Since z (0) ≤ 0, we have z ′ (t) ≤ 0 according to (48). ok
−1, then the UAV and the obstacle are in the case shown in
Consequently, z (t) ≤ z ′ (t) ≤ 0 according to the comparison
Figure 2(b). Thus,
lemma [41], namely kx (t)k ≤ ymax . 1 T
In the following, the conclusion (27) will be shown. Let ṽ ṽo = rv .
l2 o
z , x − y. Then (25) can be transformed as
  Consequently, δ (t) = r2 . Furthermore, if ξ̃o (t) >
1
ż (t) = −k (t) z (t) + k (t) ẏ .
p
2 2
r + rv and kp̃o (0)k > r, then kp̃o (t)k > r, where t > 0.
k (t)
(ii) Proof of necessity. Given any ǫo > 0, we will show if
1
If kz (0)k ≤ kmin vymax , then 2
ξ̃o (t) = r2 + rv2 − ǫo ,
1
kz (t)k ≤ vy (49) and kp̃o (0)k = r, then there exists a case that kp̃o (t)k < r,
kmin max
v T vo
where t ≥ 0. Consider a case kvkkv ok
= −1. Then the UAV
where the conclusion (26) is utilized. The equation (25) is and the obstacle are in the case shown in Figure 2(b). Thus,
further written as
T 1 T
δ (t) = ξ̃ o ξ̃o − ṽ ṽo = r2 − ǫo .
ẋ (t) = −k (t) z (t) . l2 o
According to (52), we have
It can be further written as
Z t
2
kẋ (t)k ≤ |k (t)| kz (t)k p (t) = r − e−l(t−s) lǫo ds
0
≤ kmax kz (t)k . < r2 .

Using (49) will lead to the conclusion (27).  Therefore, kp̃o (t)k < r, where t ≥ 0. 
 16

D. Proof of Theorem 1 E. Proof of Proposition 3

Proof of Conclusion (i). According to (19) in the proof of Since eo in (22) can be written as
Lemma 1, if and only if
eo = ξ̃o + (λo − εo + ε) (56)
eTo ėo keo k=rs +ro ≥0 (53) where
λo , ξ o − ξ̄ o .
then (24) holds with r̂s = rs for any keo (0)k ≥ rs + ro . The
derivative of eo is Taking the norm on both sides of (56) results in

ėo = ξ̇ − ξ̂˙o + ε̇. keo k ≤ ξ˜o + kλo k + kεo k + kεk .

Then, the inequality (53) is rewritten as If (34) holds, then

˙ ξ̃o (t) ≥ r + re − (kλo k + kεo k + kεk)

 
eTo ξ̇ − eTo ξ̂ o + eTo ε̇ ≥ 0. (54)
keo k=rs +ro
θm T s
≥r+ vo + vo τdm − kλo k (57)
i) Proof of Sufficiency. Since kε̇k ≤ vb , we have 1 − θm
where Assumptions 1-2 are utilized. The left work is to study
eTo ξ̇ − eTo ξ̂˙o + eTo ε̇
 
keo k=rs +ro
kλo k . The derivative of λo is
T ˙ 1−θ 1−θ
 
T
≥ eo ξ̇ − eo ξ̂ o − keo k kε̇k λ̇o (t) = ξ̇ o (t) + ξ̄ (t) − ξ (t − τd )
keo k=rs +ro θTs o θTs o
T ˙
 
T
≥ eo ξ̇ − eo ξ̂ o − (rs + ro ) vb . 1−θ
keo k=rs +ro =− λo (t)
θTs
 
Therefore, if (31) holds, then (54) is satisfied. ii) Proof of 1−θ θTs
+ (ξ o (t) − ξ o (t − τd )) + ξ̇ o (t) .
Necessity. The necessary condition is proved by contradiction. θTs 1−θ
Suppose (31) is not satisfied, namely there exists an e∗o with (58)
ǫ > 0 such that As for term ξ o (t) − ξ o (t − τd ) , by the mean value theorem,
˙ we have
 
e∗o T ξ̇ − e∗T
o ξ̂ o = (rs + ro ) vb − ǫ.
ke∗
o k=rs +ro
ξ o (t) − ξ o (t − τd ) = ξ̇ o (st + (1 − s) (t − τd )) τd , s ∈ [0, 1] .
e∗
Then, choose ε̇ = −vb rs +ro
o
when eo = e∗o ,
which satisfies Then
kε̇k ≤ vb . As a result, at eo = e∗o , (54) becomes kξ o (t − τd ) − ξo (t)k ≤ vo τd
e∗o
(rs + ro ) vb − ǫ − vb e∗T
o ≥0 where max ξ̇o ≤ vo is utilized. Similarly, kλo (0)k ≤ vo τd
rs + ro
by Assumption 2. Furthermore, based on the equation (58),
namely, according to Lemma 2, we have
−ǫ ≥ 0. θTs
kλo (t)k ≤ vo τd + vo
This is a contradiction. This is implies that (54) and then (53) 1−θ
θm T s
will be violated. ≤ vo τdm + vo . (59)
1 − θm
Proof of Conclusion (ii). Under Assumption 3, according to
Lemma 1, we have Using (57) and (59) yields (35). 

eTo c (t, eo ) keo k=rs +ro

≥ (rs + ro ) vm . (55)
R EFERENCES
Therefore, (31) becomes (32).
Proof of Conclusion (iii). In view of (9), according to [1] K. Balakrishnan, J. Polastre, J. Mooberry, R. Golding, P.
Sachs, “Premiering a future blueprint for our sky”[online],
Lemma 2, we have Available:https://storage.googleapis.com/blueprint/Airbus UTM Blueprint.pdf
[2] NASA, “UAS traffic management (UTM)”[online],
ξ̄˙o ≤ ξ̇o (t − τd ) ≤ vo . Available:https://utm.arc.nasa.gov
[3] SESAR, “European drones outlook 2016”[online],
Available:https://www.sesarju.eu/sites/default/files/documents/reports/European Dron
Then [4] D. Depoorter and W. Kellerer, “Designing the Air–Ground Data Links
for Future Air Traffic Control Communications”, IEEE Transactions on
eTo ξ̂˙o
 
≤ keo k ξ¯˙o + keo k kεo k Aerospace and Electronic Systems, vol. 55, no. 1, pp. 135-146, 2019.
keo k=rs +ro keo k=rs +ro [5] RTCA, “Final report of the RTCA task force 3, free flight implementa-
≤ (rs + ro ) (vo + vbo ) . tion”, Washington, DC:RTCA, Inc., 1995.
[6] J. Kosecka, C. Tomlin, G. Pappas, S. Sastry, “Generation of conflict
resolution manoeuvres for air traffic management”, 1997 IEEE/RSJ
If (33) holds, then (32) holds. Therefore, (24) holds with r̂s = International Conference on Intelligent Robot and Systems, Grenoble,
rs .  France, vol. 3, pp. 1598-1603, 1997.

[7] ICAO, “The use of displayed ADS-B data for a collision avoidance
capability in unmanned aircraft system”, UASSG/9-SN, no. 5, 2012.
[8] A. Chakrabarty, C.A. Ippolito, J. Baculi, K.S. Krishnakumar, S. Hening,
“Vehicle to vehicle (V2V) communication for collision avoidance for
multicopters flying in UTM-TCL4”, AIAA Scitech 2019 Forum, San
Diego, California, AIAA 2019-0690, 2019.
[9] S. Hayat, E. Yanmaz, R. Muzaffar, “Survey on unmanned aerial vehicle
networks for civil applications: a communications viewpoint”, IEEE [32] M. Gharibi, R. Boutaba, S.L. Waslander, “Internet of drones”, IEEE
Communications Surveys & Tutorials, vol. 18, no. 4, pp. 2624-2661,
2016.
[10] X. Dong, B. Yu, Z. Shi, Y. Zhong, “Time-varying formation control for
unmanned aerial vehicles: theories and applications”, IEEE Transactions
on Control Systems Technology, vol. 23, no. 1, pp. 340-348, 2015.
[11] J. Seo, Y. Kim, S. Kim, A. Tsourdos, “Collision Avoidance Strategies
for Unmanned Aerial Vehicles in Formation Flight”, IEEE Transactions
on Aerospace and Electronic Systems, vol. 53, no. 6, pp. 2718-2734,
2017.
[12] R. Dutta, L. Sun, D. Pack, “A decentralized formation and network
connectivity tracking controller for multiple unmanned systems”, IEEE
Transactions on Control Systems Technology, vol. 26, no. 6, pp. 2206-
2213, 2018.
[13] GSMA, “Mobile-enabled unmanned aircraft”[online],
Available:https://www.gsma.com/iot/wp-content/uploads/2018/02/Mobile-Enabled-Unmanned-Aircraft-web.pdf
[14] L.J. Glaab, C.V. Dolph, S.D. Young, N.C. Coffey, R.G. McSwain,
M.J. Logan, D.E. Harper, “Small unmanned aerial system (UAS) flight
testing of enabling vehicle technologies for the UAS traffic management
project”, NASA/TM–2018-219816, 2018.
[15] E.J. Rodrı̈¿œguez-Seda, D.M. Stipanović, M.W. Spong, “Guaranteed
collision avoidance for autonomous systems with acceleration con-
straints and sensing uncertainties”, Journal of Optimization Theory and
Applications, vol. 168, no. 3, pp. 1014-1038, 2016.
[16] X. Wang, V. Yadav, S. N. Balakrishnan, “Cooperative UAV formation
flying with obstacle/collision avoidance”, IEEE Transactions on Control
Systems Technology, vol. 15, no. 4, pp. 672-679, 2007.
[17] J. Haugen, L. Imsland, “Monitoring moving objects using aerial mobile
sensors”, IEEE Transactions on Control Systems Technology, vol. 24,
no. 2, pp. 475-486, 2016.
[18] S. Huang, R.S.H. Teo, K.K. Tan, “Collision avoidance of multi-
unmanned aerial vehicles: a review”, Annual Reviews in Control, vol.
48, pp. 147-164, 2019.
[19] A. Mcfayden and L. Mejias, “A survey of autonomous vision-based
see and avoid for unmanned aircraft systems”, Progress in Aerospace
Science, vol. 80, pp. 1-17, 2016.
[20] E. Mueller and M. J. Kochenderfer, “Simulation comparison of collision
avoidance algorithms for small multi-rotor aircraft”, AIAA Modeling and
Simulation Technologies Conference, Washington, D.C., AIAA-2016-
3674, 2016.
[21] H. Bai, D. Hsu, M.J. Kochenderfer, W.S. Lee, “Unmanned aircraft
collision avoidance using continuous-state POMDPs”, Proceedings of
Robotics: Science and Systems, Los Angeles, CA, Bai-RSS-11, 2011.
[22] H. Yang, J. Lim, S. Yoon, “Anytime RRBT for handling uncertainty
and dynamic objects”, 2016 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), Daejeon, pp. 4786-4793, 2016.
[23] X. Zhou, X. Yu, X. Peng, “UAV Collision Avoidance Based on Vary-
ing Cells Strategy”, IEEE Transactions on Aerospace and Electronic
Systems, vol. 55, no. 4, pp. 1743-1755, 2019.
[24] S.H. Arul, A.J. Sathyamoorthy, S. Patel, M. Otte, H. Xu, M.C. Lin, and
D. Manocha, “LSwarm: efficient collision avoidance for large swarms
with coverage constraints in complex urban scenes”, IEEE Robotics and
Automation Letters, vol. 4, no. 4, pp. 3940-3947, 2019.
[25] G. Angeris, K. Shah, M. Schwager, “Fast reciprocal collision avoid-
ance under measurement uncertainty”, arXiv preprint arXiv:1905.12875,
2019.
[26] H. Zhu and J. Alonso-Mora, “Chance-constrained collision avoidance
for UAVs in dynamic environments”, IEEE Robotics and Automation
Letters, vol. 4, no. 2, pp. 776-783, 2019.
[27] J. Alonso-Mora, P. Beardsley, R. Siegwart, “Cooperative collision avoid-
ance for nonholonomic robots”, IEEE Transactions on Robotics, vol. 34,
no. 2, pp. 404-420, 2018.
[28] M.H. Yamchi, R.M. Esfanjani. “Distributed predictive formation control
of networked mobile robots subject to communication delay”, Robotics
and Autonomous Systems, vol. 91, pp. 194-207, 2017.
[29] C. Virı̈¿œgh, M. Nagy, C. Gershenson, G. Vı̈¿œsı̈¿œrhelyi, “Self-
organized UAV traffic in realistic environments”, 2016 IEEE/RSJ Inter-
national Conference on Intelligent Robots and Systems (IROS), Daejeon,
pp. 1645-1652, 2016.
17
[30] J. Hoekstra, R. van Gent, R. Ruigrok, “Conceptual design of free
flight with airborne separation assurance”, Navigation, and Control
Conference and Exhibit, Boston, MA, AIAA-98-4239, 1998.
[31] S.M. Iranmanesh, E. Moradi-Pari, Y.P. Fallah, S. Das, M. Rizwan, “Ro-
bustness of cooperative forward collision warning systems to communi-
cation uncertainty”, 2016 Annual IEEE Systems Conference (SysCon),
Orlando, FL, pp. 1-7, 2016.
Access, vol. 4, pp. 1148-1162, 2016.
[33] J. Van Den Berg, S. J. Guy, M. Lin, D. Manocha, “Reciprocal n-body
collision avoidance”, in Robotics Research, Berlin, Germany:Springer,
pp. 3-19, 2011.
[34] D. Fox, W. Burgard, S. Thrun, “The dynamic window approach to
collision avoidance”, IEEE Robotics & Automation Magazine, vol. 4,
no. 1, pp. 23-33, 1997.
[35] A. Richards, T. Schouwenaars, J. P. How, E. Feron, “Spacecraft tra-
jectory planning with avoidance constraints using mixed-integer linear
programming”, Journal of Guidance Control and Dynamics, Vol. 25,
no. 4, 2002.
[36] R. Vidal, O. Shakernia, H.J. Kim, D.H. Shim, S. Sastry, “Probabilistic
pursuit-evasion games: theory, implementation, and experimental evalu-
ation”, IEEE Transactions on Robotics and Automation, vol. 18, no. 5,
pp. 662-669, 2002.
[37] A. Kamimura and T. Ohira, “Group chase and escape”, New Journal of
Physics, vol. 12, pp. 1-13, 2010.
[38] K. Oh, M. Park, H. Ahn, “A survey of multi-agent formation control”,
Automatica, 2015.
[39] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks
of agents with switching topology and time-delays”, IEEE Transactions
on Automatic Control, vol. 49, no. 9, pp. 1520–1533, 2004.
[40] B. T. Ingersoll, J. K. Ingersoll, P. DeFranco, and A. Ning, “UAV path-
planning using Bezier curves and a receding horizon approach,” AIAA
Modeling and Simulation Technologies Conference. pp. 3675, 2016.
[41] H.K. Khalil, Nonlinear Systems. Prentice-Hall: Upper Saddle River, NJ,
2002.
Quan Quan received the B.S. and Ph.D. degrees
in control science and engineering from Beihang
University, Beijing, China, in 2004 and 2010, re-
spectively. He has been an Associate Professor with
Beihang University since 2013, where he is cur-
rently with the School of Automation Science and
Electrical Engineering. His research interests include
reliable flight control, vision-based navigation, repet-
itive learning control, and time-delay systems.
Rao Fu received the B.S. degree in control sci-
ence and engineering from Beihang University, Bei-
jing, China, in 2017. He is working toward to the
Ph.D. degree at the School of Automation Science
and Electrical Engineering, Beihang University (for-
merly Beijing University of Aeronautics and Astro-
nautics), Beijing, China. His main research interests
include UAV traffic control and swarm.
澳
 18

Kai-Yuan Cai Kai-Yuan Cai received the B.S., M.S.,

and Ph.D. degrees in control science and engineer-
ing from Beihang University (Beijing University of
Aeronautics and Astronautics), Beijing, China, in
1984, 1987, and 1991, respectively. He has been a
Full Professor with Beihang University since 1995.
He is a Cheung Kong Scholar (Chair Professor),
appointed by the Ministry of Education of China in
1999. His main research interests include software
testing, software reliability, reliable flight control,
and software cybernetics.