IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO.

7, JULY 2022 3655

High-Order Control Barrier Functions

Wei Xiao , Member, IEEE, and Calin Belta , Fellow, IEEE

Abstract—We approach the problem of stabilizing a dynamical the cost is quadratic in controls. The time domain is discretized, and the
system while optimizing a cost and satisfying safety constraints state is assumed to be constant in each time interval (at its value at the
and control limitations. For (nonlinear) affine control systems and
quadratic costs, it has been shown that control barrier functions beginning of the interval). The original optimal control problem (OCP)
(CBFs) guaranteeing safety and control Lyapunov functions (CLFs) is reduced to a (possibly large) number of quadratic programs (QPs)—
enforcing convergence can be used to (conservatively) reduce one for each interval [16]. This approach is related to and arguably
the optimal control problem to a sequence of quadratic programs computationally more efficient than traditional model predictive control
(QPs). Existing works in this category have two main limitations.
First, with one exception, they are based on the assumption that
(MPC) [17].
the relative degree of the system with respect to a function en- While this approach provides a good compromise between the com-
forcing a safety constraint is one. Second, the QPs can easily putational effort necessary to compute a solution and its optimality [2],
become infeasible, in particular for problems with many safety it has two main limitations. First, it is based on the assumption that
constraints and tight control limitations. We propose high-order the relative degree of the system with respect to the function enforcing
CBFs (HOCBFs), which can accommodate systems of arbitrary
relative degrees. For each safety constraint, by using Lyapunov- the safety constraints is one. A backstepping approach was introduced
like conditions, we construct a set of controls that renders the in [18], and it was shown to work for relative degree two. A CBF
intersection of a set of sets forward invariant, which implies the method for position-based constraints with relative degree two was also
satisfaction of the original constraint. We formulate optimal control proposed in [19]. A more general form, which works for arbitrarily high
problems with constraints given by HOCBF and CLF, and pro-
relative degree constraints, was proposed in [3] and [20]. The method
pose two methods—the penalty method and the parameterization
method—to address the feasibility problem. Finally, we show how in [3] employs input–output linearization and finds a pole placement
our methodology can be extended for safe navigation in unknown controller with negative poles to stabilize the BF to zero. The resulting
environments with long-term feasibility. We illustrate the proposed BF is exponential. The authors in [21] proposed an approach to define
framework on adaptive cruise control and robot control problems. another function that is with relative degree one from the original high
Index Terms—Lyapunov methods, safety-critical control. relative degree constraint. This approach does not include all the states
in the definition of a CBF, which may decrease the problem feasibility.
Second, the QPs mentioned earlier could be infeasible, i.e., it is
I. INTRODUCTION hard to find valid CBFs that do not conflict with the control bounds,
The problem of driving a dynamical system to a desired configuration in particular for problems with many safety constraints. For the ACC
while minimizing its control effort and satisfying safety constraints and problem defined in [2], the minimum braking distance is used to
control limitations received a lot of attention in recent years [2]–[5]. simplify the process of finding a valid CBF, and these results in an
Recent works propose the use of control barrier functions (CBFs) [2] to additional complex constraint. However, this conflict is hard to address
enforce safety and control Lyapunov functions (CLFs) [6]–[8] to ensure for high-dimensional systems. The approach in [20] tried to address
convergence to desired states. conflict between CBF constraints using control-sharing BFs, without
Barrier functions (BFs) are Lyapunov-like functions [9], whose use considering the control bounds.
can be traced back to optimization problems [10]. More recently, they In this article, we define a novel notion of high-order CBF (HOCBF),
have been employed in verification and control, e.g., to prove set which is simpler and more general than the one from Nguyen and
invariance [11]–[14] and for multiobjective control [15]. CBFs are Sreenath [3]. Our HOCBFs are not restricted to exponential functions,
extensions of BFs for control systems. There are many versions of and are determined by a set of class K functions for high relative degree
CBF in the literature. The CBF defined in [2] is allowed to decrease constraints. As a generalization of the main result from Ames et al. [2], a
when far away from the boundary of the set. Simpler versions of CBFs, safety set is guaranteed to be forward invariant if the HOCBF constraint
which can approach zero inside the corresponding sets, were proposed is satisfied. In order to find a valid HOCBF, we exploit the definitions
in [4] and [5]. Time-varying CBFs were defined and used to enforce of the class K functions, and develop a methodology, called the penalty
the satisfaction of signal temporal logic (STL) formulas in [5]. method, to guarantee the feasibility of the QPs. We also propose a
Most of the works using the CBF-CLF approach are based on the parameterization method to deal with the feasibility problem when the
assumption that the (nonlinear) control system is affine in controls and penalty method fails. We provide a framework to control a system to
safely navigate in an unknown environment while ensuring long-term
feasibility.
Manuscript received December 15, 2020; revised May 13, 2021; ac- We illustrate the proposed method on adaptive cruise control (ACC)
cepted August 9, 2021. Date of publication August 18, 2021; date of
current version June 30, 2022. This work was supported in part by
and robot control problems. The simulations show the effectiveness of
National Science Foundation (NSF) under Grant IIS-1723995 and Grant the proposed HOCBF method with feasibility guarantee. For the robot
IIS-2024606. (Corresponding author: Wei Xiao.) problem, we illustrate the feasibility of the solution for unknown envi-
The authors are with the Division of Systems Engineering, Boston ronments cluttered with obstacles of similar shape but different sizes.
University, Brookline, MA 02446 USA (e-mail: [email protected]; This article is a significant extension of our recent conference
[email protected]).
Color versions of one or more figures in this article are available at paper [1]. Specifically, in addition to including the technical details
https://doi.org/10.1109/TAC.2021.3105491. related to the general definition of an HOCBF, here we introduce the
Digital Object Identifier 10.1109/TAC.2021.3105491 penalty and the parameterization methods. The penalty method simply

0018-9286 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 3656 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022

adds weights (penalties) for the class K functions in the definition of if there exist constants c1 > 0, c2 > 0, c3 > 0, and c1 ||x||2 ≤ V (x) ≤
the HOCBF from [1]. We provide conditions that guarantee problem c2 ||x||2 such that, for ∀x ∈ Rn
feasibility for this method. When these conditions are not satisfied, we
apply the alternative parameterization method, in which the class K inf [Lf V (x) + Lg V (x)u + c3 V (x)] ≤ 0. (6)
u∈U
functions are power functions multiplied by penalties. We determine
penalties and powers in the HOCBF such that the QPs are always fea- Definition 6 (Relative degree [22]): The relative degree of a (suffi-
sible and that minimize the HOCBF value for which the corresponding ciently) differentiable function b : Rn → R with respect to system (2)
constraint becomes active. This can improve the problem feasibility in is the number of times we need to differentiate it along the dynamics
an unknown environment, and we illustrate it in robot navigation. of (2) until the control u explicitly shows.
In this article, since function b is used to define a constraint b(x) ≥ 0,
we will also refer to the relative degree of b as the relative degree of the
II. PRELIMINARIES
constraint.
Definition 1 (Class K function [22]): A Lipschitz continuous func- Many existing works [2], [3], [5] combine CBF and CLF with
tion α : [0, a) → [0, ∞), a > 0 is said to belong to class K if it is strictly quadratic costs to form optimization problems. The CLF constraint
increasing and α(0) = 0. is always slacked (i.e., a slack variable is added to relax the constraint,
Lemma 1 (Lemma 4.4 in [22] and Lemma 2 in [4]): Let b : [t0 , tf ] → and minimized by adding it to the cost) when combined with CBF
R be a continuously differentiable function. If ḃ(t) ≥ −α(b(t)) ∀t ∈ to make the problem feasible; however, state convergence may not
[t0 , tf ], where α is a class K function of its argument, and b(t0 ) ≥ 0, be guaranteed. Time is discretized, and an optimization problem with
then b(t) ≥ 0 ∀t ∈ [t0 , tf ]. constraints given by CBF and CLF is solved at each time step. The
Consider a system of the form resulting problem is a sequence of QPs. The control from solving the
QP is held constant and is applied at the current time step. The dynamics
ẋ = f (x) (1)
(2) is updated, and the procedure is repeated. It is important to note that
with x ∈ X ∈ R (X denotes a closed-state constraint set) and f :
n this method works conditioned upon the fact that the control input shows
Rn → Rn globally Lipschitz. Solutions x(t) of (1), starting at x(t0 ), up in (5), i.e., Lg b(x) = 0, ∃x ∈ X and the QPs are all feasible.
t ≥ t0 , are forward complete.
We also consider affine control systems in the form III. HIGH-ORDER CBFS
ẋ = f (x) + g(x)u (2) In this section, we define high-order BFs (HOBFs) and HOCBFs.
Example: ACC: Consider the ACC problem [2] with vehicle dynam-
where g : R → Rn n×q
is globally Lipschitz, and u ∈ U ⊂ R (U q
ics
denotes a closed-control constraint set). Solutions x(t) of (2), starting
at x(t0 ), t ≥ t0 , are forward complete. v̇(t) = u(t), ż(t) = v0 − v(t) (7)
Definition 2 (Forward invariant set): A set C ⊂ Rn is forward
invariant for system (1) [or (2)] if its solutions starting at any x(t0 ) ∈ C where v(t) denotes the velocity of the ego vehicle along its lane, z(t)
satisfy x(t) ∈ C for ∀t ≥ t0 . denotes the distance between the ego and the preceding vehicles, v0 > 0
For a continuously differentiable function b : Rn → R, let denotes the speed of the preceding vehicle, and u(t) is the control input
of the ego vehicle.
C := {x ∈ Rn : b(x) ≥ 0}. (3) We require that the distance z(t) between the ego vehicle and its
Definition 3 (BF [2], [4], [5]): The function b : R → R is a n immediately preceding vehicle be greater than a constant δ > 0 for all
candidate BF for system (1) if there exists a class K function α such the times, i.e.,
that
z(t) ≥ δ ∀t ≥ t0 . (8)
ḃ(x) + α(b(x)) ≥ 0 ∀x ∈ C. (4)
Let x(t) := (v(t), z(t)) and b(x(t)) = z(t) − δ. With α(·) in Def-
Theorem 1: (see[4] and [5]) Given a set C as in (3), if there exist a inition 4 chosen as the identity function, according to (5), in order to
BF b : C → R, then C is forward invariant for system (1). ensure safety, we need to have
Definition 4 (CBF [2], [4], [5]): Given a set C as in (3), b(x) is a
candidate CBF for system (2) if there exists a class K function α s.t. v0 − v(t) + 0
 ×u(t) + z(t) − δ ≥ 0. (9)
     
Lf b(x(t)) Lg b(x(t)) b(x(t))
sup [Lf b(x) + Lg b(x)u + α(b(x))] ≥ 0 ∀x ∈ C (5)
u∈U
Note that Lg b(x(t)) = 0 in (9), i.e., u(t) does not show up. There-
where Lf and Lg denote the Lie derivatives1 along f and g, respectively. fore, we cannot use this CBF to formulate an optimization problem, as
We call the CBF in Definition 4 a candidate because α(·) is not fixed. described at the end of Section II.
In this article, we show how to define a class K function α(·) such that
there exists a control u ∈ U that satisfies (5). A CBF is completely A. High-Order BF
defined when α(·) is specified. This applies to Definition 3 as well.
Theorem 2 (see[4] and [5]): Given a CBF b with the associated As in [5], we consider a time-varying function to define an invariant
set C from (3), any Lipschitz continuous controller u(t) ∀t ≥ t0 that set for system (1). For a mth-order differentiable function b : Rn ×
satisfies (5) renders the set C forward invariant for (2). [t0 , ∞) → R, we define a sequence of functions ψi : Rn × [t0 , ∞) →
Definition 5 (CLF [8]): A continuously differentiable function V : R, i ∈ {1, . . . , m} in the form
Rn → R is a globally and exponentially stabilizing CLF for system (2)
ψi (x, t) = ψ̇i−1 (x, t)+αi (ψi−1 (x, t)), i ∈ {1, . . . , m} (10)
1 The Lie derivative of a function along a vector field captures the change in where αi (·), i ∈ {1, . . . , m} denote class K functions of their argument
the value of the function along the vector field (see, e.g., [22]). and ψ0 (x, t) = b(x, t).

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022 3657

We further define a sequence of sets Ci (t), i ∈ {1, . . . , m} associ- Given an HOCBF b, we define the set of control that satisfies
ated with (10) in the form

Ci (t) = {x ∈ Rn : ψi−1 (x, t) ≥ 0}, i ∈ {1, . . . , m}. (11) Khocbf (x, t) = {u ∈ U : Lm

f b(x, t)+Lg Lf
m−1
b(x, t)u
∂ m b(x, t)
Definition 7: Let Ci (t), i ∈ {1, . . . , m} be defined by (11) and + + O(b(x, t)) + αm (ψm−1 (x, t)) ≥ 0}.
ψi (x, t), i ∈ {1, . . . , m} be defined by (10). A function b : Rn × ∂tm
(14)
[t0 , ∞) → R is a candidate HOBF for (1) if it is mth-order differen-
tiable and there exist differentiable class K functions αi , i ∈ {1,. . .m}
s.t. Theorem 4: Given an HOCBF b(x, t) from Definition 8 with the
associated sets Ci (t), i ∈ {1, . . . , m} defined by (11), if x(t0 ) ∈
ψm (x, t) ≥ 0 (12) C1 (t0 )∩, . . . , ∩Cm (t0 ), then any Lipschitz continuous controller
u(t) ∈ Khocbf (x(t), t) ∀t ≥ t0 renders the set C1 (t)∩, . . . , ∩Cm (t)
for all (x, t) ∈ C1 (t)∩, . . . , ∩Cm (t) × [t0 , ∞). forward invariant for system (2).
Similar to Definition 4, an HOBF is defined when αi (·), i ∈ Proof: Since u(t) is Lipschitz continuous and u(t) only shows up
{1, . . . , m} are found. in the last equation of (10) when we take Lie derivative on (10), we
Theorem 3: The set C1 (t)∩, . . . , ∩Cm (t) is forward invariant for have that ψm (x, t) is also Lipschitz continuous. The system states in
system (1) if b(x, t) is an HOBF. (2) are all continuously differentiable, so ψi (x, t), i ∈ {1, . . . , m} are
Proof: If b(x(t), t) is an HOBF, then ψm (x(t), t) ≥ 0 for also continuously differentiable. Therefore, the HOCBF has the same
∀t ∈ [t0 , ∞), i.e., ψ̇m−1 (x(t), t) + αm (ψm−1 (x(t), t)) ≥ 0. By property as the HOBF in Definition 7, and the proof is the same as
Lemma 1, since x(t0 ) ∈ Cm (t0 ) (i.e., ψm−1 (x(t0 ), t0 )) ≥ 0, and Theorem 3. 
ψm−1 (x(t), t) is an explicit form of ψm−1 (t)), then ψm−1 (x(t), t)) ≥ Remark 2: The general, time-varying HOCBF introduced in Defi-
0 ∀t ∈ [t0 , ∞), i.e., ψ̇m−2 (x(t), t) + αm−1 (ψm−2 (x(t), t)) ≥ 0. nition 8, can be used for general, time-varying constraints (e.g., STL
Again, by Lemma 1, since x(t0 ) ∈ Cm−1 (t0 ), we also have specifications [5]) and systems. However, many problems, such as the
ψm−2 (x(t), t)) ≥ 0 ∀t ∈ [t0 , ∞). Iteratively, we can get ACC and robot control problems that we consider in this article, have
x(t) ∈ Ci (t) ∀i ∈ {1, . . . , m} ∀t ∈ [t0 , ∞). Therefore, the set time-invariant system dynamics and constraints. Therefore, in the rest
C1 (t)∩, . . . , ∩Cm (t) is forward invariant.  of this article, we focus on time-invariant versions for simplicity.
Remark 1: The sets Ci (t), i ∈ {1, . . . , m} should have a nonempty Remark 3 (Relationship between time-invariant HOCBF and ex-
intersection at t0 in order to satisfy the forward invariance con- ponential CBF in [3]): In Definition 7, if we set class K functions
dition starting from t0 in Theorem 3. If b(x(t0 ), t0 ) ≥ 0, we α1 , α2 . . . αm to be linear functions with positive coefficients, then we
can always choose proper class K functions αi (·), i ∈ {1, . . . , m} can get exactly the same formulation as in [3] that is obtained through
to make ψi (x(t0 ), t0 ) ≥ 0 ∀i ∈ {1, . . . , m − 1}. There are some input–output linearization, i.e.,
extreme cases, however, when this is not possible. For exam-
ple, if ψ0 (x(t0 ), t0 ) = 0 and ψ̇0 (x(t0 ), t0 ) < 0, then ψ1 (x(t0 ), t0 )
is always negative no matter how we choose α1 (·). Similarly, ψi (x) = ψ̇i−1 (x) + ki ψi−1 (x), i ∈ {1, . . . , m} (15)
if ψ0 (x(t0 ), t0 ) = 0, ψ̇0 (x(t0 ), t0 ) = 0, and ψ̇1 (x(t0 ), t0 ) < 0,
ψ2 (x(t0 ), t0 ) is also always negative, etc. To deal with such extreme where ki > 0, i ∈ {1, . . . , m}. Therefore, the time-invariant version
cases (as with the case when b(x(t0 ), t0 ) < 0), we would need a HOCBF defined in this article is a generalization of the exponential
feasibility enforcement method, which is beyond the scope of this CBF introduced in [3].
article. Remark 4 (Comparison between HOCBF and MPC): In an MPC
approach [17], the optimization is defined over a receding horizon.
B. High-Order CBF Compared to the myopic HOCBF method considered in this article
(i.e., the optimization is over one step), the MPC optimization is more
Definition 8: Let Ci (t), i ∈ {1, . . . , m} be defined by (11) and
likely to be feasible. However, it is more difficult, as it is, in general
ψi (x, t), i ∈ {1, . . . , m} be defined by (10). A function b : Rn ×
a nonlinear program. On the other hand, the HOCBF approach can
[t0 , ∞) → R is a candidate HOCBF of relative degree m for system
handle nonlinear (affine) dynamics, and the corresponding optimization
(2) if there exist differentiable class K functions αi , i ∈ {1, . . . , m} s.t.
problems are easy to solve. The myopia of an HOCBF is significantly
∂ m b(x, t) improved if a valid HOCBF is found offline, as shown in Section IV-A.
sup [Lm
f b(x, t) + Lg Lf
m−1
b(x, t)u + This offline computation is hard to be performed in an MPC approach.
u∈U ∂tm
Example revisited: For the ACC problem introduced at the beginning
+O(b(x, t)) + αm (ψm−1 (x, t))] ≥ 0 (13) of Section III, the relative degree of the constraint from (8) is 2.
Therefore, we need an HOCBF with m = 2. We choose quadratic class
for all (x, t) ∈ C1 (t)∩, . . . , ∩Cm (t) × [t0 , ∞). Lf and Lg denote the K functions for both α1 (·) and α2 (·), i.e., α1 (b(x(t))) = b2 (x(t)) and
partial Lie derivatives w.r.t. x along f and g, respectively. α2 (ψ1 (x(t))) = ψ12 (x(t)). In order for b(x(t)) := z(t) − δ to be an
Similar to Definition 4, an HOCBF is defined when αi (·), i ∈ HOCBF for (7), a control input u(t) should satisfy
{1, . . . , m} are specified. In this article, we show how to find such
functions. In the aforementioned equation, O(·) is given by
L2f b(x(t)) + Lg Lf b(x(t))u(t) + 2b(x(t))Lf b(x(t))

m−1
O(b(x, t)) = Lif (αm−i ◦ ψm−i−1 )(x, t) +(Lf b(x(t)))2 + 2b2 (x(t))Lf b(x(t)) + b4 (x(t)) ≥ 0. (16)
i=1

∂ i (αm−i ◦ ψm−i−1 )(x, t) Note that Lg Lf b(x(t)) = 0 in (16) and the initial conditions are
+ .
∂ti b(x(t0 )) ≥ 0 and ḃ(x(t0 )) + b2 (x(t0 )) ≥ 0.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 3658 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022

IV. OPTIMAL CONTROL WITH TIME-INVARIANT HOCBF unsafe sets [24], in which case the proposed approximation method
in [23] can still work.
In this section, we show how to find a valid HOCBF. Consider an
OCP for system (2) with the cost defined as
 tf A. Penalty Method
J(u(t)) = C(||u(t)||)dt (17) In (10), we multiply the class K function αi (·) with penalties
t0 (weights) pi ≥ 0, i ∈ {1, . . . , m} in the form
where || · || denotes the two-norm of a vector. t0 and tf denote the ψi (x) = ψ̇i−1 (x) + pi αi (ψi−1 (x)), i ∈ {1, . . . , m}. (21)
initial and final times, respectively, and C(·) is a strictly increasing
function of its argument (such as the energy consumption function The sets X and U are closed. Let
C(||u(t)||) = ||u(t)||2 ). Assume a time-invariant (safety) constraint
Umin := inf [−Lg Lm−1 b(x(t))u]
b(x) ≥ 0 with relative degree m has to be satisfied by system (2). Then, x∈X,u∈U
f

the control input u should satisfy the time-invariant HOCBF version

of the constraint from (13)
Umax := sup [−Lg Lm−1
f b(x(t))u]
f b(x) + Lg Lf b(x)u + O(b(x)) + αm (ψm−1 (x)) ≥ 0 (18)
Lm m−1 x∈X,u∈U

for all x ∈ C1 ∩, . . . , ∩Cm (Ci , i ∈ {1, . . . , m} denotes  the

time-invariant version of Ci (t)), where O(b(x)) = m−1 Fmin := inf [Lm
f b(x)].
i=1 x∈X
Lif (αm−i ◦ ψm−i−1 )(x).
The following theorem provides conditions for the feasibility guar-
If convergence to a given state is required in addition to optimality
antee of the QP (20).
and safety, then, as in [2], HOCBF can be combined with CLF. Suppose
Theorem 5: If Umax ≤ Fmin , then there exist (small enough) pi ≥
the control bound U for (2) is defined as
0, i ∈ {1, . . . , m} such that the control limitations (19) do not conflict
U = {u ∈ Rq : umin ≤ u(t) ≤ umax ∀t ∈ [t0 , tf ]} (19) with the HOCBF constraint (18) ∀x(t0 ) ∈ C1 ∩ · · · ∩ Cm .
Proof: It follows from the sequence of equations in (21) that
where umin , umax ∈ Rq . In order to solve this optimization problem, p1 , p2 , . . . , pm−1 will appear in all terms of O(b(x)) in (18), i.e.,
we use the QP-based approach (suppose C(||u(t)||) = ||u(t)||2 ) in- O(b(x)) = 0 if pi = 0 ∀i ∈ 1, 2, . . . , m − 1. Since pm shows up in
troduced at the end of Section II, i.e., we partition the time interval the last equation of (21), we have that
[t0 , tf ] into a set of equal time intervals {[t0 , t0 + Δt), [t0 + Δt, t0 +
2Δt), . . . }, where Δt > 0. In each interval [t0 + ωΔt, t0 + (ω + −Lg Lm−1
f b(x)u ≤ Lm
f b(x)

1)Δt) (ω = 0, 1, 2, . . . ), we keep the state constant at its value at the pi = 0 ∀i ∈ 1, 2, . . . , m.

if Since Lmf b(x) ≥ Fmin , if
beginning of the interval and also assume the control is constant, and re-
−Lg Lf b(x)u ≤ Fmin , then the last constraint is satisfied.
m−1
formulate the optimization problem as a sequence of QPs. Specifically,
The control bound on u in (19) always satisfies
at t = t0 + ωΔt (ω = 0, 1, 2, . . . ), we solve
Umin ≤ −Lg Lm−1 b(x)u ≤ Umax .
(u∗ (t), δ ∗ (t)) = arg min ||u(t)||2 + pδ 2 (t) f
u(t),δ(t)
Since Fmin ≥ Umax , the intersection of the sets determined by the last
s.t.(18), (19) and two inequalities is always nonempty, i.e., the intersection of the control
bounds (19) and the HOCBF constraint (18) is always nonempty.
Lf V (x) + Lg V (x)u + c3 V (x) ≤ δ (20) We conclude that there exist small enough penalties p1 ≥ 0, p2 ≥
where δ is a slack variable used to relax (soften) the CLF constraint 0, . . . , pm ≥ 0 such that the control limitations (19) will not conflict
and p > 0 is a weight. After solving (20), we integrate (2) with with the HOCBF constraint (18). 
control u∗ (t) kept constant during [t0 + ωΔt, t0 + (ω + 1)Δt). This The following corollary provides simpler conditions for systems
QP-based method is suboptimal compared with the original OCP (17), [such as (7)] that satisfy extra properties.
as the optimizations are performed pointwise. Corollary 1: If 0 ∈ U and Lm f b(x) ≥ 0 ∀x ∈ X, then there exist

Constraint (18) may conflict with (19), in which case we cannot find (small enough) pi ≥ 0, i ∈ {1, . . . , m} such that the control limitations
a valid HOCBF. If this happens, the OCP becomes infeasible. In the (19) do not conflict with the HOCBF constraint (18) ∀x(t0 ) ∈ C1 ∩
rest of this section, we propose a two-stage methodology to find a valid · · · ∩ Cm .
HOCBF, which is based on offline computations of solutions to (20). We Proof: Similar to the proof of the last theorem, we have
first apply the penalty method (see Section IV-A,) assuming the class −Lg Lm−1 b(x)u ≤ Lm
f f b(x)
K functions are given. If this fails, i.e., the conditions under which the
penalty method works are not satisfied, we turn to the parameterization if pi = 0 ∀i ∈ 1, 2, . . . , m. Since Lm f b(x) ≥ 0 ∀x ∈ X, if
method (see Section IV-B). We find a valid HOCBF based on the worst- −Lg Lm−1f b(x)u ≤ 0, then the last constraint is satisfied. The 0
case initial state for some symmetric unsafe sets (such as circular unsafe vector is included in the last equation, and 0 ∈ U . Therefore, there
sets). For such sets, the problem feasibility does not heavily depend exist small enough penalties p1 ≥ 0, p2 ≥ 0, . . . , pm ≥ 0 such
on the initial state and the worst-case initial condition is also easy to that the control limitations (19) do not conflict with the HOCBF
find. For example, for a spherical obstacle, the worst-case initial state constraint (18). 
corresponds to maximum velocity directed at the center of the sphere. Example revisited: For the ACC problem introduced in Section III,
With some conservatism, other geometries can also be dealt with by L2f b(x) = 0. If 0 is included in the control bound, then from Corollary
covering them with symmetric sets—in our recent work [23], we used 1, it follows that HOCBF constraints do not conflict with the control
disks. Note that a valid HOCBF might be hard to find for nonconvex bound when we choose small enough penalties p1 , p2 for α1 (·), α2 (·).

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022 3659

Remark 5 (Applying the penalty method): Given the class K func-

tions αi (·), i ∈ {1, . . . , m} in an HOCBF b(x), if the QP (20) becomes
infeasible for some x ∈ C1 ∩ · · · ∩ Cm , or it becomes infeasible at
some time t ∈ [t0 , tf ], we restart from time t0 and add penalties to the
class K functions as in (21). Note that, when penalties are applied, the
sets Ci , i = 2, . . . , m will be affected. By random selection, we try to
find values for the penalties such that x(t0 ) ∈ C1 ∩ · · · ∩ Cm . If the
optimization problem becomes feasible, then we are done. Otherwise,
we decrease the value of p1 , as p1 shows up in all the ψi (·) functions
in (21), and thus decreasing p1 is the most efficient way among all the
penalties to make the problem feasible. However, decreasing p1 can
significantly shrink C2 , which might result in x(t0 ) ∈ / C2 , as shown
in (11). In order to avoid this, we can proceed with decreasing p2 , and
recursively try to find penalties such that x(t0 ) ∈ C1 ∩ · · · ∩ Cm . If
we are not successful, then we will turn to the parameterization method
described next.

B. Parameterization Method
When the conditions in Theorem 5 or Corollary 1 are not satisfied
or no penalties can be found because of x(t0 ) as in Remark 5, we use
the parameterization method, in which we also determine the class K
functions. Since power functions are mostly used as class K functions,
we can explicitly write (21) as
qi
Remark 7 (Comparison between the penalty and parameterization
ψi (x) = ψ̇i−1 (x) + pi ψi−1 (x), i ∈ {1, . . . , m} (22) methods): The penalty method is simpler and more intuitive than
the parameterization method as we can just choose small penalties
where qi ≥ 1 ∀i ∈ {1, . . . , m}. The penalties pi , i ∈ {1, . . . , m} and
to make the problem feasible (see Remark 5). The parameterization
powers qi , i ∈ {1, . . . , m} are parameters of the HOCBF, and they
method does not necessarily choose small pi ’s, but searches through
determine at what time the HOCBF constraint (18) becomes active
the whole parameter space, and thus is less conservative and more
(i.e., it is satisfied as an equality). If the HOCBF constraint becomes
computationally expensive. The parameterization method works well
active when system (2) is close to the obstacle (i.e., b(x) is close to
for unknown environments as we make the HOCBF constraints active
0), system (2) may require a large control input such that the safety
as late as possible to allow the system to detect obstacles.
constraint b(x) ≥ 0 could be enforced by the HOCBF, which could
Remark 8 (The effect of time discretization): In order to implement
possibly conflict with the control bound U . Thus, the OCP can become
the HOCBF method in real systems, we need to discretize the time,
infeasible. Ideally, we would like to choose the parameters such that
as described at the end of Section II. The forward invariance of the
(20) is feasible in [t0 , tf ]. We may just randomly sample the penalties
safety sets might not be guaranteed in between the discretization time
and powers such that the problem (20) is feasible [assume there exists
instants. The work in [21] addresses this issued by trying to find closed-
a feasible solution for the problem and there exist such parameters
form solutions to the problem. The sampling approach from Cortez
that can make the problem feasible; otherwise, we need to consider
et al. [25] is used to ensure the constraint satisfaction in the time
all possible class K functions instead of just power functions in (22)].
intervals between the discretization instants. The self-triggered method
However, we do not want the HOCBF constraint (18) to be active when
from Yang et al. [26] can also be used to determine discretization times
system (2) is far from the corresponding obstacle (i.e., b(x) is large)
ensuring invariance in continuous time.
as the obstacles may not be detected before the HOCBF constraint
Remark 9 (The use of extended class K functions): We can define
(18) becomes active in an unknown environment. If this happens, the
αi (·), i ∈ {1, . . . , m} in Definition 8 as extended class K functions
initial conditions of an HOCBF in Theorem 4 may not be satisfied
(α : [−a, a] → [−∞, ∞] as in Definition 1) to ensure robustness of
or the HOCBF constraint may conflict with the control bound U , and
an HOCBF to perturbations [21]. However, the use of extended class
thus, the safety is not guaranteed. Therefore, we want to choose the
K functions cannot ensure a constraint to be satisfied if it is initially
parameters such that the value of the HOCBF b(x) when the constraint
violated for a relative degree one CBF, which can also cause a similar
first becomes active is minimized.
problem in an HOCBF since ψi (x) in (10) is recursively defined.
Remark 6 (Applying the parameterization method): We can use
a gradient-descent method to find the optimal penalties and powers
such that the problem is feasible while also minimizing b(x(ta )), V. CASE STUDIES AND RESULTS
where ta ∈ [t0 , tf ] is the time at which the HOCBF constraint first In this section, we complete the ACC case study and introduce a robot
becomes active under the worst-initial condition (such as maximum control problem. All the computations and simulations were conducted
approaching speed). We randomly sample for pi , qi , and start with any in MATLAB.
pi , qi , i ∈ {1, . . . , m} such that the QPs are all feasible in [t0 , tf ] and
take b(x(ta )) as the objective function to minimize with pi , qi , i ∈ A. Adaptive Cruise Control
{1, . . . , m} as decision variables. We evaluate the gradient of b(x(ta ))
with respect to pi , qi , then minus pi , qi by the gradient times a learning For the dynamics given by (7), we consider a cost J(u) =
tf
rate γ > 0 and solve the QP (20) again ∀t ∈ [t0 , tf ] to find a smaller t0
u2 (t)dt, and we require the vehicle to achieve a desired speed
possible b(x(ta )). We repeat this process for each feasible sample (see vd = 24 m/s. For this, we define a CLF V (x) = (v − vd )2 (see Defi-
Algorithm 1). nition 5).

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 3660 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022

heading angle, v denotes the linear speed, and u1 , u2 are the two
control inputs (turning speed and forward acceleration). Note that the
dynamics are in the form (2), with x = (x, y, θ, v)T , u = (u1 , u2 )T ,
f = (v cos θ, v sin θ, 0, 0)T , and g = (0, 0; 0, 0; 1, 0; 0, 1).
In this problem, we have two objectives: (o1) minimize energy
t
consumption J(u(t)) = t0f (u21 (t) + u22 (t))dt, and (o2) reach desti-
nation (xd , yd ) ∈ R2 , during time interval [t1 , t2 ], t0 ≤ t1 ≤ t2 ≤ tf ,
and two constraints: (c1) safety (x(t) − xo )2 + (y(t) − yo )2 ≥ r2 ,
where (xo , yo ) ∈ R2 denotes the location of a circular obstacle and
r = 7 m is its size (a little larger than the actual size, which is
6 m), and (c2) robot limitationsvmin ≤ v(t) ≤ vmax , u1,min ≤ u1 (t) ≤
u1,max , u2,min ≤ u2 (t) ≤ u2,max , where vmin = 0 m/s, vmax = 2 m/s,
u1,max = −u1,min = 0.2 rad/s, and u2,max = −u2,min = 0.5 m/s2 .
Fig. 1. Control input u(t) as b(x(t)) → 0 for different p when using We use HOCBFs to (strictly) impose constraints (c1) and
quadratic class K function. All the solid lines (black, red, blue) start from
b(x) = 90. They coincide before the corresponding HOCBF constraint (c2) and two CLFs V1 (x) = (θ − atan( xyd −y ))2 , V2 (x) = (v −
d −x
2
becomes active (e.g., the red solid line can only be seen after b(x) <= vd ) , vd = 2 m/s to achieve objective (o2). We capture objective (o1)
45), when the solid line starts overlapping with its associated dashed in the cost of the optimization problem. For constraint (c1), we define
line. The arrows denote the changing trend for b(x(t)) with respect to
an HOCBF b(x) = (x(t) − xo )2 + (y(t) − yo )2 − r2 with m = 2.
time.
Since L2f b(x) = 2v 2 is guaranteed to be nonnegative, the penalty
method always works given proper x(t0 ). However, we use the pa-
rameterization method for (c1), since we also wish to minimize the
HOCBF value when the HOCBF constraint first becomes active. This
approach gives good results in an unknown environment with obstacles,
as shown below. We use two HOCBFs to impose the speed part of (c2):
bmax (x) = vmax − v and bmin (x) = v − vmin . Both have relative degree
1, and Lf bmax (x) = Lf bmin (x) = 0. If we set u = 0, the speed will not
change. Therefore, these HOCBF constraints do not conflict with the
control bound, and we do not need to use the penalty or parameterization
methods.
We find penalties p1 , p2 and powers q1 , q2 for a worst-case scenario.
We consider maximum initial speed v(t0 ) = vmax and initial position
Fig. 2. Variations of functions b(x(t)) and ψ1 (x(t)) for linear (p = 1)
(x(t0 ), y(t0 )) = (5, 25 m). We assume the obstacle center (xo , yo ) =
and quadratic (p = 0.02) class K functions, respectively. b(x(t)) ≥ 0 and (32, 25 m) and the destination (xd , yd ) = (45, 25 m) are aligned, and
ψ1 (x(t)) ≥ 0 imply the forward invariance of C1 ∩ C2 . the initial heading θ(t0 ) = 0 is also parallel to this line [see Fig. 3(a)].
We study the feasibility robustness of the solution (i.e., how feasibility
is affected by changes in state and/or environment).
When the destination component yd is exactly 25 m, the robot stops
We consider a control constraint −0.4g ≤ u(t) ≤ 0.4 g, g =
before the obstacle (if p1 , p2 , q1 , q2 are feasible for the QP), i.e., it can-
9.81 m/s2 . The relative degree of (8) is two, and we define three
not arrive at the destination. We call this stop point an equilibrium point,
different HOCBFs as in Definition 8 by choosing square root, linear and
as shown in Fig. 3(a). However, if yd has a positive offset (arbitrary
quadratic class K functions for both α1 (·), α2 (·) in (21) with penalties
small), the robot can overpass the obstacle and arrive at the destination
p1 = p2 = p > 0. The other parameters are the same as in [1].
following the left trajectories shown in Fig. 3(a). Otherwise, the robot
We define b(x) = z − δ as an HOCBF with m = 2. Since L2f b(x) =
will produce right trajectories also shown in Fig. 3(a). Therefore, we
0, the conditions in Corollary 1 are satisfied and we can find small
choose a small offset for yd (i.e., yd = 25.0000001 m) when trying to
enough p1 , p2 such that the problem is feasible. We present the penalty
find the optimal p1 , p2 , q1 , q2 .
case study for quadratic class K functions in Fig. 1. The dashed lines
We randomly sample p1 , p2 , q1 , q2 (p1 , p2 ∈ (0, 3], q1 , q2 ∈ [1, 3])
denote the values of the right-hand side of the HOCBF constraint (i.e.,
Lm b(x)+O(b(x))+αm (ψm−1 (x)) to get 2000 points, and run simulations for 30 s. We use the algorithm in
f
−Lg Lm−1 b(x)
), and the solid lines are the optimal Remark 6 to optimize each sample and get the optimal (p∗1 , p∗2 , q1∗ , q2∗ ) =
f
controls. When the dashed lines and solid lines coincide, the HOCBF (0.7535, 0.6664, 1.0046, 1.0267) such that the QPs are feasible and the
constraint for b(x) is active. HOCBF value is minimized when the HOCBF constraint first becomes
In Fig. 1, the HOCBF constraint does not conflict with the braking active. The HOCBF constraint is active when b(x) = Dmin (Dmin =
limitation −cd g when p = 0.02 for a quadratic class K function. The 5.4 m2 , a distance metric instead of the real distance). In the rest of this
minimum control input (negative) increases as p decreases. Then, we set section, we study the feasibility robustness of the proposed method,
p to be 1, 0.02 for linear and quadratic class K functions, respectively. assuming the optimal (p∗1 , p∗2 , q1∗ , q2∗ ) given above. Note that the QP
We present the forward invariance of the set C1 ∩ C2 , where C1 := feasibility does not depend on the specific initial condition as long as
{x(t) : b(x(t)) ≥ 0} and C2 := {x(t) : ψ1 (x(t)) ≥ 0} in Fig. 2. the robot initially has a distance (in terms of b(x)) of at least Dmin
from the obstacle, and the initial state is within the predefined bound
(see [27]).
B. Robot Control
1) Feasibility Robustness to the Heading Angle: In this
Consider the unicycle model for a wheeled mobile robot ẋ = case, we only change the value of destination component yd . Based
v cos θ, ẏ = v sin θ, v̇ = u2 , θ̇ = u1 ,x, y denote the location, θ is the on yd = 25.0000001 m and yd = 24.9999999 m, we further offset yd

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022 3661

Fig. 3. Robot control problem using HOCBFs and the parameterization method. (a) Trajectories under different parameters. (b) Trajectories under
different obstacle-approaching angles. (c) Trajectories under different obstacle-approaching speeds. (d) Trajectories for obstacles of different sizes.

arrive sequentially at points a := (39, 35 m), b := (30, 15 m), c :=

(38, 40 m), and d := (20, 28 m). The robot can safely arrive at these
four destinations with the penalties and powers (p1 , p2 , q1 , q2 ) =
(0.7535, 0.6664, 1.0046, 1.0267) as calculated earlier, which shows
good feasibility robustness. The robot trajectory is shown in Fig. 4.
The computation time to solve the QP at each time step is less than
0.01 s (Intel(R) Core(TM) i7-8700 CPU @ 3.2 GHz×2).

VI. CONCLUSION
We extended BFs and CBFs to HOBFs and HOCBFs, and showed
how they can be used to solve OCPs with safety requirements and
control limitations for systems with high relative degree. We showed
how the new definitions can be used to significantly increase the
feasibility of the OCPs. We applied the proposed framework to an ACC
Fig. 4. Safe exploration in an unknown environment.
problem and to a robot navigating in an unknown environment facing
real-time safety constraints. In the future, we will investigate the use of
machine learning techniques to improve system performance.
by +2 m for yd = 25.0000001 m (−2 m for yd = 24.9999999 m), and
generate 7 destinations for both cases, respectively. These 14 destina-
tions are all feasible, which shows good feasibility robustness of the REFERENCES
penalty method to changes in the heading angle when approaching the [1] W. Xiao and C. Belta, “Control barrier functions for systems with high
obstacle, as shown in Fig. 3(b). The HOCBF values when the HOCBF relative degree,” in Proc. 58th IEEE Conf. Decis. Control, Nice, France,
constraint becomes active are all smaller than Dmin . 2019, pp. 474–479.
2) Feasibility Robustness to the Approaching Speed: [2] A. D. Ames, J. W. Grizzle, and P. Tabuada, “Control barrier function based
quadratic programs with application to adaptive cruise control,” in Proc.
We vary the approaching speed to the obstacle between 1.8 and 2.5 m/s 53rd IEEE Conf. Decis. Control, 2014, pp. 6271–6278.
(2 m/s was the value for the original problem). All these values are [3] Q. Nguyen and K. Sreenath, “Exponential control barrier functions for
all feasible, which shows good feasibility robustness of the penalty enforcing high relative-degree safety-critical constraints,” in Proc. Amer.
method to the change in speed when approaching the obstacle. The Control Conf., 2016, pp. 322–328.
[4] P. Glotfelter, J. Cortes, and M. Egerstedt, “Nonsmooth barrier functions
HOCBF values when the HOCBF constraint becomes active increase
with applications to multi-robot systems,” IEEE Control Syst. Lett., vol. 1,
as the approaching speed increases, as shown in Fig. 3(c). no. 2, pp. 310–315, Oct. 2017.
3) Feasibility Robustness to the Obstacle Size: Here, we [5] L. Lindemann and D. V. Dimarogonas, “Control barrier functions for signal
only change the obstacle size from the predefined value r = 7 m. We temporal logic tasks,” IEEE Control Syst. Lett., vol. 3, no. 1, pp. 96–101,
consider a range of r between 2 and 9 m. The results show good Jan. 2019.
[6] E. Sontag, “A Lyapunov-like stabilization of asymptotic controllability,”
feasibility robustness to the change of obstacle size as the QPs are SIAM J. Control Optim., vol. 21, no. 3, pp. 462–471, 1983.
always feasible and the robot can safely arrive at its destination, as [7] Z. Artstein, “Stabilization with relaxed controls,” Nonlinear Anal.: Theory,
shown in Fig. 3(d). The HOCBF values when the HOCBF constraint Methods, Appl., vol. 7, no. 11, pp. 1163–1173, 1983.
becomes active do not change under different-size obstacles. [8] A. D. Ames, K. Galloway, and J. W. Grizzle, “Control Lyapunov functions
and hybrid zero dynamics,” in Proc. 51st IEEE Conf. Decis. Control, 2012,
Finally, in order to show the feasibility robustness is independent of
pp. 6837–6842.
the location of the obstacles, we present an application of robot safe [9] P. Wieland and F. Allgower, “Constructive safety using control barrier
exploration in an unknown environment. Suppose the robot is equipped functions,” in Proc. 7th IFAC Symp. Nonlinear Control Syst., 2007,
with a sensor [ 23 π field of view and 7 m (greater than the one corre- pp. 473–478.
sponding to Dmin ) sensing distance with 1 − m sensing uncertainty] to [10] S. P. Boyd and L. Vandenberghe, Convex Optimization. New York, NY,
USA: Cambridge Univ. Press, 2004.
detect the obstacles, and there are three unknown obstacles (to the robot) [11] K. P. Tee, S. S. Ge, and E. H. Tay, “Barrier Lyapunov functions for the
whose center locations are (32, 25 m), (28, 35 m), and (30, 40 m) control of output-constrained nonlinear systems,” Automatica, vol. 45,
with radius 6, 5, and 6 m, respectively. The robot is required to no. 4, pp. 918–927, 2009.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.
 3662 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 7, JULY 2022

[12] J. P. Aubin, Viability Theory. Berlin, Germany: Springer, 2009. [20] X. Xu, “Constrained control of input-output linearizable systems using
[13] S. Prajna, A. Jadbabaie, and G. J. Pappas, “A framework for worst-case control sharing barrier functions,” Automatica, vol. 87, pp. 195–201, 2018.
and stochastic safety verification using barrier certificates,” IEEE Trans. [21] A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function
Autom. Control, vol. 52, no. 8, pp. 1415–1428, Aug. 2007. based quadratic programs for safety critical systems,” IEEE Trans. Autom.
[14] R. Wisniewski and C. Sloth, “Converse barrier certificate theorem,” in Control, vol. 62, no. 8, pp. 3861–3876, Aug. 2017.
Proc. 52nd IEEE Conf. Decis. Control, Florence, Italy, 2013, pp. 4713– [22] H. K. Khalil, Nonlinear Systems, 3rd ed. Englewood Cliffs, NJ, USA:
4718. Prentice-Hall, 2002.
[15] D. Panagou, D. M. Stipanovic, and P. G. Voulgaris, “Multi-objective [23] W. Xiao et al., “Rule-based optimal control for autonomous driving,” in
control for multi-agent systems using Lyapunov-like barrier functions,” Proc. ACM/IEEE 12th Int. Conf. Cyber-Phys. Syst., 2021, pp. 143–154.
in Proc. 52nd IEEE Conf. Decis. Control, Florence, Italy, 2013, pp. 1478– [24] A. Robey et al., “Learning control barrier functions from expert demon-
1483. strations,” in Proc. 59th IEEE Conf. Decis. Control, 2020, pp. 3717–3724.
[16] K. Galloway, K. Sreenath, A. D. Ames, and J. Grizzle, “Torque saturation [25] W. S. Cortez, D. Oetomo, C. Manzie, and P. Choong, “Control barrier
in bipedal robotic walking through control Lyapunov function based functions for mechanical systems: Theory and application to robotic
quadratic programs,” IEEE, vol. 3, pp. 323–332, 2015. grasping,” IEEE Trans. Control Syst. Technol., vol. 29, no. 2, pp. 530–545,
[17] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Control: Mar. 2021.
Theory, Computation, and Design. Santa Barbara, CA, USA: Nob Hill [26] G. Yang, C. Belta, and R. Tron, “Self-triggered control for safety critical
Publishing, 2017. systems using control barrier functions,” in Proc. Amer. Control Conf.,
[18] S.-C. Hsu, X. Xu, and A. D. Ames, “Control barrier function based 2019, pp. 4454–4459.
quadratic programs with application to bipedal robotic walking,” in Proc. [27] W. Xiao, C. Belta, and C. G. Cassandras, “Feasibility guided learning for
Amer. Control Conf., 2015, pp. 4542–4548. constrained optimal control problems,” in Proc. 59th IEEE Conf. Decis.
[19] G. Wu and K. Sreenath, “Safety-critical and constrained geometric control Control, 2020, pp. 1896–1901.
synthesis using control Lyapunov and control barrier functions for systems
evolving on manifolds,” in Proc. Amer. Control Conf., 2015, pp. 2038–
2044.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on April 08,2024 at 09:14:49 UTC from IEEE Xplore. Restrictions apply.