Uncertainty Quantification for Recursive Estimation

in Adaptive Safety-Critical Control

Max H. Cohen1 , Makai Mann2 , Kevin Leahy3 , and Calin Belta4

Abstract— In this paper, we present a framework for online a barrier function [5] provides a methodology to certify
parameter estimation and uncertainty quantification in the set invariance without computing a system’s reachable set.
context of adaptive safety-critical control. The key insight Such ideas naturally extend to control systems via control
enabling our approach is that the parameter estimate gen-
erated by the continuous-time recursive least squares (RLS) barrier functions (CBFs) [4], [5] that allow for the design of
arXiv:2304.01901v2 [eess.SY] 12 Mar 2024

algorithm at any point in time is an affine transformation controllers that enforce set invariance.
of the initial parameter estimate. This property allows for Given the duality between Lyapunov and barrier functions
parameterizing such estimates using objects that are closed [4], recent works have begun to transfer Lyapunov-based
under affine transformation, such as zonotopes, and enables adaptive control techniques to a barrier function setting.
the efficient propagation of such set-based estimates as time
progresses. We illustrate how such an approach facilitates For example, [6] introduced the notion of an adaptive
the synthesis of safety-critical controllers for systems with CBF (aCBF) that allows for synthesizing controllers and
parametric uncertainty and additive disturbances using control parameter update laws enforcing set invariance conditions for
barrier functions, and demonstrate the utility of our approach nonlinear systems with uncertain parameters. A key insight
through illustrative examples. of follow-up works [3], [7]–[13] is that directly extending
I. I NTRODUCTION classical Lyapunov-based techniques to a barrier function
setting can lead to overly conservative controllers, and that
As autonomous systems are deployed in complex en- such conservatism can be reduced by taking a robust adaptive
vironments, it is essential for such systems to adapt to control approach. Here, one assumes known bounds on the
uncertainties that may be challenging to characterize until system parameters, designs a robust controller enforcing
deployment. Recently, there has been a surge of interest in safety for all possible realizations of the parameters, and then
using data-driven and learning-based techniques to endow reduces the robustness of this controller as the uncertainty
such systems with adaptive capabilities. This integration of in the parameter estimates shrinks.
learning and control, however, raises questions regarding The success of the aforementioned robust adaptive ap-
the reliability and safety of such learning-enabled systems. proaches relies on various uncertainty quantification mech-
In a control-theoretic context, simultaneous learning and anisms that associate an uncertainty metric to a given pa-
control is the focus of adaptive control [1]–[3] where the rameter estimate. For aCBFs, such an idea was introduced
primary objective is stabilization of dynamical systems with in [7], where uncertainty quantification is achieved using
parametric uncertainties. set membership identification (SMID) – a technique com-
Although adaptive control schemes with stability guar- monly used in adaptive model predictive control [14]–[16]
antees have existed for decades, there has been much less that maintains a set-based parameter estimate, rather than a
work on extending such ideas to address other important pointwise estimate as in classical adaptive control [1]. Such
properties, such as safety. The property of safety – informally set-based estimates are generally computed by solving an
thought of as requiring a system to “never do anything bad” optimization problem at each time instant and often allow
– has often been formalized using the notion of set invariance one to account for additive disturbances in addition to un-
[4]. In an analogous fashion to how Lyapunov functions certain parameters [16]. One limitation of SMID approaches
are leveraged to certify the stability of equilibrium points is that the number of constraints in the optimization problem
without integrating a system’s vector fields, the concept of may grow unbounded for extended time horizons, requiring
1 Department of Mechanical and Civil Engineering, California Institute various complexity reduction techniques to facilitate real-
of Technology, Pasadena, CA 91125; [email protected] time implementation [14]. Alternate uncertainty quantifica-
2 Massachusetts Institute of Technology Lincoln Laboratory, Lexington,
tion techniques used for aCBFs include concurrent learning
MA 02421; [email protected] [17], [18] – an adaptive control method that facilitates
3 Department of Robotics Engineering, Worcester Polytechnic Institute,
Worcester, MA 01609; [email protected] parameter convergence under relaxed excitation conditions.
4 Department of Electrical and Computer Engineering, University of Such conditions are related to the rank of a data matrix that is
Maryland, College Park, MD 20742; [email protected] updated as time progresses, and allows for bounding the rate
This work is supported by the NSF under grants DGE-1840990 and
IIS-2024606. Any opinions, findings, conclusions or recommendations ex- of convergence of the parameter estimates using Lyapunov-
pressed in this material are those of the author(s) and do not necessarily like arguments [3], [8]–[10]. Such an approach allows for
reflect the views of the NSF. The NASA University Leadership initiative efficiently computing bounds on the parameter estimation
(grant #80NSSC20M0163) provided funds to assist the authors with their
research, but this article solely reflects the opinions and conclusions of its error; however, the bounds are often conservative, especially
authors and not any NASA entity. in the presence of additive disturbances.
 In this paper, we unite SMID and concurrent learning
<latexit sha1_base64="gXCYFDdP+kY2xlB6hpAoBcKzSuU=">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</latexit>

1.5 (a)
adaptive control for online parameter estimation and uncer- = 0.1, w(t) = 0.0
tainty quantification in the context of adaptive safety-critical = 0.1, w(t) = 0.5
control. Our starting point is the classic recursive least 1 = 0.001, w(t) = 0.5
squares (RLS) algorithm for online parameter estimation.

k✓(t)k
Inspired by [17], we first demonstrate how incorporating a Without H
<latexit sha1_base64="vE2h9JPS5JHDkq9DkTuA9o4413E=">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</latexit>

˜
history stack of data into RLS facilitates parameter con- 0.5

<latexit sha1_base64="2EKoQspXEabFep74H2PqwE55XX4=">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</latexit>
vergence under milder conditions than the traditional PE With H
<latexit sha1_base64="iEihuV960j/+J70kECGB0rxy6ho=">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</latexit>

condition, and analyze the impact of additive disturbances

on such convergence results (Sec. III). We then demonstrate 0
how the pointwise parameter estimates produced by RLS can 80 0 <latexit sha1_base64="bWNMfx54XHIfjF1wzmG5O5akV3w=">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</latexit>

20 40 60
(b)

(t)
be directly used to construct a set-based parameter estimate, 40
allowing one to quantify the uncertainty in the original 0

<latexit sha1_base64="D7HuG8/gOFuMxAAhQ48w/dWM3IQ=">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</latexit>
pointwise estimates. Our key insight is that the parameter 0 20 40 60
t
<latexit sha1_base64="/EXVznWUhYIzIoiPeDG9UtuTRp0=">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</latexit>

estimate produced by RLS at any given point in time is an Fig. 1. Evolution of the parameter estimation errors (a) and value of γ (b)
from (6) corresponding to Example 1. The solid curves correspond to the
affine transformation of the initial estimate (Sec. IV). This estimates using H and the dashed curves to those that do not exploit H.
observation allows for parameterizing such estimates using
various objects that are closed under affine transformation, R, feature vector ϕ(t) ∈ Rp , and a disturbance input w(t) ∈
such as zonotopes [19]. By replacing the initial pointwise R. Throughout this paper, we make the standing assumption
parameter estimate with a zonotope containing the true pa- that all signals in (1) are bounded. The primary objective
rameters, this approach allows to compute, at each instant in in online parameter estimation is to build an estimate θ̂ :
time, another zonotope containing the true parameters using R≥0 → Rp of θ completely online based upon knowledge
only the initial zonotope and pointwise estimate produced by of the signals y(·) and ϕ(·). Most often, this is accomplished
RLS. Ultimately, this results in a recursive SMID technique by solving the optimization problem:
that maintains the computational efficiently of RLS while
min J(θ̂, t), (2)
producing set-based parameter estimates that allow for uncer- θ̂∈Rp
tainty quantification in the presence of additive disturbances.
where J : Rp × R≥0 → R is a cost function, continuously
We demonstrate how this approach facilitates the synthesis
differentiable in its first argument and piecewise continuous
of safety-critical controllers for uncertain nonlinear systems
in its second, that represents a metric between observed
(Sec. V) and illustrate the efficacy of our approach with
measurements y(t) and predicted measurements ϕ(t)⊤ θ̂,
various examples.
allowing to construct the parameter estimate at time t as:
II. P RELIMINARIES AND P ROBLEM F ORMULATION
θ̂(t) = arg min J(θ̂, t). (3)
Notation: For a matrix A ∈ Rn×n the notation A ≥ 0 (A > θ̂∈Rp
0) denotes that A is positive semidefinite (positive definite).
Rather than explicitly solving (2) for every t, one may re-
The symbol I denotes the identity matrix. Given matrices
cursively implement the solution (3) to optimization problem
A, B with the same number of rows, we define hcat(A, B) ˙
as their horizontal concatenation. We use ∥ · ∥p to denote the (2) by propagating the ordinary differential equation θ̂ =
(induced) p-norm for a (matrix) vector. For a continuously τ (θ̂, t), where τ : Rp × R≥0 → Rp is a vector field, locally
differentiable function h : Rn → R and a vector field f : Lipschitz in its first argument and piecewise continuous in its
Rn → Rn , we define Lf h(x) := ∇h(x) · f (x) as the Lie second, from an initial condition θ̂ 0 ∈ Rp corresponding to
derivative of h along f . the best guess of θ. This allows for efficiently constructing
The high-level problem considered in this paper is online a parameter estimate θ̂(t) for each t from streaming data,
parameter estimation. Here, we construct an online estimate but generally does not provide information related to the
θ̂ ∈ Rp of a fixed parameter vector θ ∈ Rp – typically uncertainty of such an estimate. In this paper, we seek to
corresponding to uncertain parameters of a control system – develop an efficient parameter estimation routine that also
based upon input-output data generated by a parameter mea- allows for quantifying the uncertainty in such an estimate,
surement model. Keeping in line with standard approaches in which we accomplish by recursively updating an entire set
adaptive control [1], [2], we assume this measurement model of possible parameters consistent with the observed data.
is affine in the parameters and of the form:
III. R ECURSIVE -BATCH L EAST S QUARES R EGRESSION
⊤
y(t) = ϕ(t) θ + w(t), (1) In this section, we present a hybrid approach that combines
where the mappings1 y : R≥0 → R, ϕ : R≥0 → Rp , and aspects of batch and recursive least squares techniques for
w : R≥0 → R associate to each t ≥ 0 a target value y(t) ∈ regression in the context of estimating the parameters θ ∈
Rp of the regression model from (1) based on ideas from
1 We assume y(t) is a scalar only for ease of exposition - the results here concurrent learning adaptive control [3], [17], [20]. Central
can be extended to vector-valued y(t) and matrix-valued ϕ(t). to such approaches is the concept of a history stack.
  Z t −1
Definition 1. A collection H(t) = {(y(tj ), ϕ(tj )}Mj=1 , Γ(t) = e−βt P + e−β(t−s) Φ(s)⊤ Φ(s)ds , (9)
where y and ϕ are as in (1) and tj ∈ R≥0 , is said to be 0
a history stack for (1). which can be recursively computing by solving:
One may think of a history stack H simply as a dataset ˙
h i
θ̂ = ΓΦ(t)⊤ Y(t) − Φ(t)θ̂ , (10)
of M ∈ N target and feature values (y(tj ), ϕ(tj )) ∈ R × Rp
sampled from (1) at various point in time tj ∈ R≥0 . Note Γ̇ = βΓ − ΓΦ(t)⊤ Φ(t)Γ, (11)
that the stack itself may vary with time as data is added or
−1
removed. Thus, H(t) returns the data present in the stack at with initial conditions θ̂(0) = θ̂ 0 and Γ(0) = P . The
time t, which need not be the same as the data present in above ODEs correspond to the continuous-time recursive
the stack at time t′ in the sense that we may have H(t) ̸= least-squares (RLS) algorithm, which can be interpreted as
H(t′ ) for t ̸= t′ . Algorithms for populating H with data are a Kalman filter for the static dynamical system θ̇ = 0 with
outlined in [18]. As originally shown in [17], constructing measurement Y(t) = Φ(t)θ + W(t), where Γ corresponds
a parameter update law using the data present in H, rather to the covariance4 of the estimate θ̂. By following the same
than only the data generated by (1) at each time, allows for steps as in the proof of [1, Corollary 4.3.2], one can show
relaxing the classical persistence of excitation (PE) condition that, provided the FE condition is satisfied, then there exist
often required for parameter convergence. To introduce these positive constants γ1 , γ2 ∈ R>0 such that:
relaxed conditions, we concatenate the data in H as:
γ1 I ≤ Γ(t) ≤ γ2 I, ∀t ∈ R≥0 , (12)
ϕ(t1 )⊤
   
y(t1 )
 y(t2 )   ϕ(t2 )⊤  where γ2 is inversely proportional to γ in (6). This implies
Y(t) :=  .  ∈ RM , Φ(t) := 
  
..  ∈ RM ×p ,
 Γ(t) is bounded and positive definite for all t ∈ R≥0 , thereby
 ..   .  implying the existence of Γ(t)−1 for all t ∈ R≥0 . Our
y(tM ) ⊤ objective now is to quantify the uncertainty in the estimate
ϕ(tM )
(4) θ̂, which is accomplished by studying the evolution of the
which satisfy: parameter estimation error:
Y(t) = Φ(t)θ + W(t), (5) θ̃(t) := θ − θ̂. (13)

where W(t) := (w(t1 ), w(t2 ), . . . , w(tM )) ∈ RM . The The following lemma provides a useful representation of θ̃.
following definition outlines a relaxed excitation condition Lemma 1. Let H satisfy the FE condition and suppose that θ̂
for parameter convergence. and Γ satisfy (8) and (9), respectively. Then, for all t ∈ R≥0 ,
Definition 2. A history stack H is said to satisfy the finite the parameter estimation error (13) satisfies:
excitation (FE) condition if there exists a time T ∈ R≥0 and θ̃(t) =e−βt Γ(t)P(θ − θ̂ 0 )
positive constant γ ∈ R>0 such that for all t ∈ [T, ∞) Z t (14)
⊤
− Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds.
Φ(t) Φ(t) ≥ γI. (6) 0

The FE condition2 requires Φ(t)⊤ Φ(t) to be invertible3 Proof. Using (5) and (8), θ̂ can be expressed as:
for all t ≥ T , thereby ensuring the existence of a solution −βt
Z t
to Y(t) = Φ(t)θ for each t ≥ T . A necessary condition for θ̂(t) =e Γ(t)Pθ̂ 0 + Γ(t) e−β(t−s) Φ(s)⊤ Φ(s)θds
0
such a solution to exist is M ≥ p. To construct a parameter Z t
update law using H, we introduce the cost function: + Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds.
0
1 t −β(t−s)
Z
J(θ̂, t) = e ∥Y(s) − Φ(s)θ̂∥22 ds Using the fact that Γ(t) is invertible for each t ∈ R≥0 , we
2 0 (7) have θ = Γ(t)Γ(t)−1 θ, which implies that
1
+ e−βt (θ̂ − θ̂ 0 )⊤ P(θ̂ − θ̂ 0 ),
Z t
2 θ = e−βt Γ(t)Pθ + Γ(t) e−β(t−s) Φ(s)⊤ Φ(s)θds.
0
where P ∈ Rp×p is positive definite and β ∈ R>0 is
a discount factor. The estimate that minimizes the above Substituting the previous two expressions into (13) then
expression is given by: yields (14), as desired.
 Z t  In the disturbance-free case (i.e., W(t) ≡ 0), the preced-
−βt −β(t−s) ⊤
θ̂(t) = Γ(t) e Pθ̂ 0 + e Φ(s) Y(s)ds , (8) ing lemma shows that θ̂ exponentially converges to θ when
0 H satisfies the FE condition. The following result provides
2 In contrast, the PE condition requires the existence of a time T > 0
a conservative bound on θ̃ when disturbances are present.
such that the matrix tt+T Φ(s)⊤ Φ(s)ds is invertible for all t ≥ 0.
R
3 The data collection algorithms in [18] ensure that λ ⊤ 4 Since θ is not viewed as a random variable, formally, Γ is not a
min (Φ(t) Φ(t))
is non-decreasing; hence, if (6) holds at a single instant in time, then it also covariance matrix; however, as demonstrated later on Γ does indeed encode
holds for all times thereafter. a measure of uncertainty associated with the estimate θ̂.
Proposition 1. Let the conditions of Lemma 1 hold and The preceding result provides a conservative bound (16)
define on the parameter estimation error in the presence of dis-
w̄ := sup ∥Φ(t)⊤ W(t)∥2 . (15) turbances, and shows that, under the FE condition, the
t≥0
parameter estimates converge to a ball of radius γ2βw̄ centered
Then, for all t ≥ 0, the parameter estimation error satisfies: at θ. Recalling that γ2 is inversely proportional to γ from (6),
γ2 w̄ this implies that the effects of disturbances can be mitigated
∥θ̃(t)∥2 ≤ γ2 e−βt ∥P(θ̂ 0 − θ)∥2 + . (16) by increasing5 γ (i.e., by collecting “better” data – see [18]
β
for data collection strategies). Although the bound in (16)
Proof. We begin by noting that:
is theoretically appealing, it is challenging to exploit online
for quantifying the uncertainty in the estimate θ̂ since it
∥θ̃(t)∥2 = e−βt Γ(t)P(θ − θ̂ 0 )
depends on γ2 , which, in turn, depends on γ, which is
Z t unknown a priori. The main ideas introduced in this section
− Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds are illustrated in the following simple example.
0 2
−βt Example 1. Consider the parameter measurement model in
= e Γ(t)P(θ̂ 0 − θ)
(1) defined by ϕ(t) = [2 sin(t/5) 12 cos(2t/5)]⊤ for some
Z t disturbance signal w(·) with parameters θ = (1, −1) ∈ R2 .
+ Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds , To estimate these parameters online, we build a history stack
0 2
H(t) by periodically sampling the signals y(t) and ϕ(t) and
which can be upper-bounded as: storing6 the values in H(t). The data in H(t) is then used
∥θ̃(t)∥2 ≤∥e−βt Γ(t)P(θ̂ 0 − θ)∥2 to define the vector Y(t) and matrix Φ(t) as in (4), which
Z t are used to estimate the parameters using the update laws
+ Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds in (10) and (11). The evolution of the resulting parameter
0 2 estimation errors for different values of β and disturbance
−βt
≤e ∥Γ(t)P(θ̂ 0 − θ)∥2 signals w(t) are illustrated by the solid lines in Fig. 1a.
Z t The value of γ satisfying the bound in (6) for each t ≥ 0
+ Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds , is displayed in Fig. 1b, indicating that the history stack
0 2
satisfies the FE condition. As guaranteed by Lemma 1 and
where the second equality follows from the homogeneity of Proposition 1, the parameter estimation error exponentially
norms, the first inequality from the triangle inequality, and converges to zero in the absence of disturbances (blue)
the second inequality from the homogeneity of norms. It then and converges to a neighborhood of zero in the presence
follows from (12), the Cauchy-Schwarz inequality, and the of disturbances (red and purple). Note that these signals
fact that the induced matrix 2-norm is monotonic that: also satisfy the PE condition, and the evolution under the
∥Γ(t)P(θ̂ 0 − θ)∥2 ≤∥Γ(t)∥2 · ∥P(θ̂ 0 − θ)∥2 traditional RLS algorithm (i.e., without the history stack [1,
Ch. 4]) is illustrated by the dashed lines in Fig. 1a. Although
≤∥γ2 I∥2 · ∥P(θ̂ 0 − θ)∥2
the estimation error converges to zero without the history
≤γ2 ∥P(θ̂ 0 − θ)∥2 . stack, the convergence is slower, the residual bounds are
Using the Cauchy-Schwarz inequality, we also have: larger, and such bounds are more sensitive to β.
Z t
η(t) = Γ(t) e−β(t−s) Φ(s)⊤ W(s)ds IV. R ECURSIVE S ET-M EMBERSHIP I DENTIFICATION
0 2
Z t
−β(t−s) ⊤
In this section we demonstrate how the estimation scheme
≤∥Γ(t)∥2 · e Φ(s) W(s)ds . introduced in the previous section naturally extends to quan-
0 2
tifying the uncertainty in the parameter estimate θ̂ through
After using (12) and the fact that the norm of an integral is
the matrix Γ. This is accomplished by blending classical pa-
upper-bounded by the integral of the norm, the above may
rameter estimation techniques from adaptive control with set
be bounded as:
Z t membership identification (SMID) techniques. Ultimately,
η(t) ≤γ2 e−β(t−s) Φ(s)⊤ W(s) ds this unification leads to the ability to efficiently propagate
2 a set of all possible parameter values using the update
Z0 t
laws proposed in the previous section. To facilitate this
≤γ2 e−β(t−s) Φ(s)⊤ W(s) ds. approach, we assume that the set of all possible parameters
0 2
can be represented as a zonotope, which are widely used for
Using (15) then allows the above to be bounded as:
reachability analysis [19], [21].
Z t
γ2 w̄   γ w̄
2
η(t) ≤γ2 w̄ e−β(t−s) ds = 1 − e−βt ≤ .
0 β β 5 Although (16) suggests increasing β may help attenuate disturbances as
well, γ2 also depends on β, and increasing β may also decrease γ2 (see
Combining the preceding inequalities yields (16), as desired. [1, Cor. 4.3.2] for more details).
6 Here, the size of H is not fixed and grows as time progresses.
 2 2 5

0 0

✓2
0
✓2

✓2
5
2 2
2 0 2 2 0 2 10 0 10
✓1 ✓1 ✓1
Fig. 2. Zonotopes generated by the RLS algorithm from Example 2. In each plot, the black dot denotes θ and the change in opacity of the zonotopes
represents the evolution of time, where more transparent shapes denote zonotopes at earlier times and less transparent shapes denote zonotopes at later
times. In each simulation, the update laws are defined by a history stack with M = 20 entries, which is filled by periodically storing values of y(t) and
ϕ(t) until the stack is full.

Definition 3. A set Z ⊂ Rp is said to be a zonotope if: The following theorem constitutes the first main result of
n X q o this paper, providing conditions under which θ is guaranteed
Z = θ ∈ Rp : θ = c + λi Gi , −1 ≤ λi ≤ 1 , (17) to remain in the zonotopes generated by (19).
i=1
Theorem 1. Let the conditions of Lemma 1 hold and define
where c ∈ R is the zonotope’s center and Gi ∈ Rp , i ∈
p

{1, . . . , q}, are its generators, which form a generator matrix w̄∞ := sup ∥Φ(t)⊤ W(t)∥∞ . (21)
t≥0
G = hcat(G1 , . . . , Gq ).
Furthermore, suppose that:
To emphasize that a zonotope depends on c and G, we
write Z(c, G) ⊂ Rp to denote a zonotope Z with center c w̄∞
∥P(θ̂ 0 − θ)∥∞ + ≤ 1. (22)
and generator matrix G. The main property of zonotopes we β
exploit is that they are closed under affine transformation: Then:
AZ(c, G) + b = Z(Ac + b, AG), (18) θ ∈ Z(θ̂(t), Γ(t)), ∀t ≥ 0. (23)
for a matrix A and vector b of appropriate dimensions. Proof. We check that Lemma 2 holds for:
The primary observation that facilitates our extension of the
previous estimation scheme to set-based estimates is that the {θ} = Z(θ, 0) ⊆ Z(θ̂(t), Γ(t)),
point-wise estimate θ̂(t) from (8) at any given point in time for each t ≥ 0. Noting that G1 = 0 and G2 = Γ, we find
is simply an affine transformation of the initial condition: that (20a) is satisfied by Q = 0 since Γ(t) is positive definite
Z t
−βt
for each t ≥ 0. Noting that c1 = θ and c2 = θ̂, we find that
θ̂(t) = e Γ(t)P θ̂ 0 + Γ(t) e−β(t−s) Φ(s)⊤ Y(s)ds . (20b) is satisfied by:
| {z } 0
A(t) | {z }
b(t) z = Γ(t)−1 (θ̂(t) − θ) = −Γ(t)−1 θ̃(t).
Replacing θ̂ 0 with Z(θ̂ 0 , P−1 ) then yields: To complete the proof, we must show that Q and z satisfy
A(t)Z(θ̂ 0 , P −1
) + b(t) =Z(A(t)θ̂ 0 + b(t), A(t)P −1
) (20c). We first note that since Q = 0 and the induced ∞-
norm for matrices is equal to the absolute row sum, we have:
=Z(θ̂(t), Γ(t)).
(19) ∥ hcat(Q, z)∥∞ = ∥z∥∞ = ∥ − Γ−1 (t)θ̃(t)∥.
Ultimately, we would like to show that θ ∈ Z(θ̂(t), Γ(t))
for each t, which allows for quantifying the uncertainty in It then follows from the above and Lemma 1 that:
a given parameter estimate θ̂(t) via Z(θ̂(t), Γ(t)). Further Z t
results are facilitated by the following lemma. ∥z∥∞ = e−βt P(θ̂ 0 − θ) + e−β(t−s) Φ(s)⊤ W(s)ds
0 ∞
Lemma 2 ([22]). Consider zonotopes Z(c1 , G1 ) ⊂ Rp and which, after using the triangle inequality and the homogene-
Z(c2 , G2 ) ⊂ Rp with centers c1 , c2 ∈ Rp and generators ity of norms, implies that:
G1 ∈ Rp×q , G2 ∈ Rp×r . Then, Z(c1 , G1 ) ⊆ Z(c2 , G2 ) if
there exists a pair (z, Q) ∈ Rp × Rr×q such that ∥z∥∞ ≤e−βt ∥P(θ̂ 0 − θ)∥∞
Z t
G1 = G2 Q (20a) + e−β(t−s) Φ(s)⊤ W(s)ds
0 ∞
c2 − c1 = G2 z (20b) Z t
≤∥P(θ̂ 0 − θ)∥∞ + e−β(t−s) Φ(s)⊤ W(s)ds ,
∥ hcat(Q, z)∥∞ ≤ 1. (20c) 0 ∞
where the second inequality follows from e−βt ≤ 1. Using 2 <latexit sha1_base64="MZyD/LnQXdunfqLgJo83oAz/5WA=">AAANCnichdfbbts2GABgtTu0S70t3S4HDMSCAinQBfaKHbCrNmma9JQ4aU5NlAUkRduMKVEVqTiO4N3tbtjDDLvYMGxXu9ob7Bn2EiNlO7+t380EBKD4kT8PJimGpUoaW6//c+36W2+/8+6Nm+/N3aq9/8GH87c/2jM6z7jY5Vrp7IBRI5RMxK6VVomDNBM0Zkrss+6K9/0zkRmpkx3bT8VxTNuJbElOrcs6mV8MY2o7nKricLAYsg61RWg7wtLBYv3uvZCt0TimLnn3ZH6hvlQvH4ITjVFi4UGw9e8Pn/663Ty5fevvMNI8j0ViuaLGHDXqqT0uaGYlV2IwF+ZGpJR3aVscuWRCY2GOi3JIA3LH5USkpTP3l1hS5k7WKGhsTD9mrqQfgKmaz5xlR7ltfXNcyCTNrUj4sKFWrojVxM8PiWQmuFV9l6A8k66vhHdoRrl1szjVio9ttVZmaihFJDJ55mb3bHqIBYvdeyJ6XLsZTaIipFk7lsmgCHUqMmp15megJ21HyVha4yaqHd7zJWbUo+f/V8+VqNQzOTt1Q7O6UrUYwfdOKlW2XVE/TMaK7aodgh1WbQNso2oxHxw1jovLdbfQqJbYGRTfhVang7m5O2RZq4gkeczK6Z8sxi5EpscNtYp6NQzTiQD2rYzDKd0TGXe7hihhLQ5MoRpFURkgQ8gBOcIIMEI40VXhMRNT3AJuDdyqJRubO6vfEre+SGjFuWUtIlukr3Pidr9yy7dH3bZxi7rn1rAgrJxFnXzuYxBfYanSfhsaaKPOdQA7CCWgRHgKeIqwC9hFqAAVwjgGjSfmw4k7K5KWktwa4jcFcacZGVUsz5NyKtqZEN3xr1+digRio83HJlYc2i4sBUwRvgZ8jTADzBAaQIPQAlqEOWCO8AzwDGEPsIfwHBAdM6wP2Ed4AXgxsSPzNL1yRz6Eag9RzGXAZYQrgCsIHwE+QrgKuIrwMeBjhGuAawjXAdcRPgF8gvAp4FOEzwCfIXwO+BzhC8AXCDcA0UnONgE3ETYBmwi3ALcQbgOirw57CfgS4Q7gDsJdwF2Ee4B7CPcB9xEeAB4gfAX4CuEh4OHMr9PUKVXd9uUlzR+F4wsbOhhoPnaKN3/akSP1KXQ00AwKjF5QBHMZwciZe/mq/pcXy1H9YRr9Xi6s0smozPgNLbPLbjZn9LF52cfmVB/dPZf466O7vRH/YTA575BFpXXXvRMlrbtIk56QWVR+RHRu/XfEDYlGp7mx/k6LPhsu5vBmY7qiV9x37fqcsnV/x0G3AW1xDZc5eGMA9ANLFYlKgGHeVY3OqFRptRrDTZu7/Teqd32c2PtiqfHV0pdb7t8AEgyfm8EnwWfBYtAIvg4eBOtBM9gNePBT8EvwR/Bn7cfaz7Xfar8Pi16/NqrzcTD11P76D1gJ/G4=</latexit>
<latexit sha1_base64="gXCYFDdP+kY2xlB6hpAoBcKzSuU=">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</latexit>

4 <latexit sha1_base64="MZyD/LnQXdunfqLgJo83oAz/5WA=">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</latexit>

ˆ (a) ˆ
<latexit sha1_base64="bWNMfx54XHIfjF1wzmG5O5akV3w=">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</latexit>

a similar argument as in the proof of Proposition 1, we may Z(✓(0), (0)) Z(✓(0), (0)) (b)
bound the second term in the above inequality as:
0 0 <latexit sha1_base64="rsCipJcqhFSvTBmwk/CCeG5kg1w=">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</latexit>

Z t ˆ ), (T ))
Z(✓(T
w̄∞
<latexit sha1_base64="rsCipJcqhFSvTBmwk/CCeG5kg1w=">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</latexit>

ˆ ), (T ))
Z(✓(T
e−β(t−s) Φ(s)⊤ W(s)ds ≤ ,
0 ∞ β 2 4
2 0 2 4 0 4
which, after using (22), implies that:
Fig. 3. Initial and final zonotopes (T = 60) corresponding to the different
w̄∞ cases from Example 2. Here, Fig. 3(a) corresponds to the zonotopes from
∥z∥∞ ≤ ∥P(θ̂ 0 − θ)∥∞ + ≤ 1. Fig. 2(a) and Fig. 3(b) corresponds to the zonotopes from Fig. 2(c).
β
It then follows that the conditions of Lemma 2 are satisfied, as illustrated in Fig. 2. To remedy this, one may define:
which implies that (23) holds, as desired. P(t) = Z(θ̂(0), Γ(0)) ∩ Z(θ̂(t), Γ(t)),
The preceding theorem provides the machinery to con- which produces a polytope (not necessarily a zonotope) that
struct a zonotope at each instant in time that is guaranteed satisfies θ ∈ P(t) for all t ∈ R≥0 .
to contain the true parameters. The pair (θ̂ 0 , P) from (7)
captures prior knowledge of the parameters in that θ̂ 0 cor- V. ROBUST A DAPTIVE S AFETY-C RITICAL C ONTROL
responds to the best initial guess of the parameters and P We now illustrate how the techniques from the previous
represents the confidence of such a guess. The condition in section can be used to develop safety-critical controllers for
(22) indicates that P must be chosen small enough for a systems with parametric uncertainty. We consider nonlinear
given level of initial parameter estimation error ∥θ − θ̂ 0 ∥∞ control systems with parametric uncertainty of the form:
and the discount factor β must be chosen large enough for a
ẋ = f (x) + F(x)θ + g(x)u + d, (24)
given level of disturbance affecting (1) so that Z(θ̂(t), Γ(t))
contains θ for all t ≥ 0. where x ∈ Rn is the state, u ∈ Rm is the control, d ∈ Rn
is a disturbance, and f : Rn → Rn , g : Rn → Rn×m , and
Example 2. We continue Example 1 to demonstrate how
F : Rn → Rn×p are locally Lipschitz continuous. Hence,
the approach in this section is used to construct a set-
given a locally Lipschitz feedback controller u = k(x, θ̂, Γ),
based estimate of θ. We begin with the disturbance-free
a piecewise continuous disturbance signal t 7→ d(t), and an
case, which, by Theorem 1, requires selecting P such that
initial condition (x0 , θ̂ 0 , Γ0 ) ∈ Rn × Rp × Rp×p , there exist
∥P(θ̂ 0 − θ)∥∞ ≤ 1. We take θ̂ 0 = 0 and assume prior
trajectories t 7→ x(t), θ̂(t), Γ(t) of the closed-loop system:
bounds on θ in the sense that ∥θ̂ 0 − θ∥ ≤ 2, allowing to
take P = 12 I to satisfy (22). The initial estimates are then ẋ =f (x) + F(x)θ + g(x)k(x, θ̂, Γ) + d(t),
used to define the initial conditions of the ODEs from (10) ˙
θ̂ =ΓΦ(t)⊤ Y(t) − Φ(t)θ̂ , (25)
 
and (11) as θ̂(0) = θ̂ 0 and Γ(0) = P−1 , respectively, the
solutions of which t 7→ θ̂(t), Γ(t) are used to construct a Γ̇ =βΓ − ΓΦ(t)⊤ Φ(t)Γ,
zonotope Z(θ̂(t), Γ(t)) whose evolution is illustrated in Fig.
defined on some maximal interval of existence I ⊆ R≥0 ,
2a. As guaranteed by Theorem 1, these zonotopes contain θ
which we take to be the nonnegative real line I = R≥0 for
for all time. We now introduce a disturbance w(t) = 0.1 and
ease of exposition. Our main objective in this section is to
propagate the ODEs in (10) and (11) with the same initial
design feedback controllers for (24) such that the closed-
conditions and discount factor β as in the disturbance-free
loop system (25) is safe, a concept often formalized using
case. The resulting zonotopes are illustrated Fig. 2b, where,
the framework of set invariance. Formally, we say that (25) is
after a certain period of time, the zonotopes no longer contain
safe on a set C ⊂ Rn from an initial condition (x0 , θ̂ 0 , Γ0 ) ∈
θ as the condition in (22) is not satisfied. To remedy this one
C × Rp × Rp×p if the resulting trajectory satisfies x(t) ∈ C
must decrease P and increase β commensurate with the value
for all t ∈ R≥0 . By considering sets of the form:
of w̄∞ . To satisfy (22) we take P = 14 I and β = 1.5 and
propagate the ODEs in (10) and (11), the results of which C = {x ∈ Rn : h(x) ≥ 0}, (26)
are shown in Fig. 2(c). With these modified values of P and
β, the zonotopes in Fig. 2c contain the true parameters at all for a continuously differentiable function h : Rn → R, one
times, but are more conservative. Although the intermediate can certify the safety of (25) using barrier functions.
zonotopes are more conservative than in the disturbance-free Definition 4. A continuously differentiable function h :
case, the zonotopes generated by such an approach for longer Rn → R defining a set C ⊂ Rn as in (26) is said to be
time horizons significantly reduce the conservatism of the a robust adaptive barrier function (RaBF) for (25) on C if
initial zonotope (cf. Fig. 3). zero is a regular value of h and there exists an extended class
K∞ function α such that for all (x, θ̂, Γ) ∈ Rn ×Rp ×Rp×p :
Remark 1. Although Theorem 1 guarantees that θ ∈
Z(θ̂(t), Γ(t)) for all t ∈ R≥0 , this does not necessarily Lf h(x) + LF h(x)θ + Lg h(x)k(x, θ̂, Γ)
(27)
imply that Z(θ̂(t), Γ(t)) ⊆ Z(θ̂(0), Γ(0)) for all t ∈ R≥0 , ¯
≥ −α(h(x)) + d∥∇h(x)∥ 2,
where d¯ := supt≥0 ∥d(t)∥2 . A RaCBF allows one to construct the set-valued map:
n
The following lemma shows that the existence of an RaBF K(x, θ̂, Γ) := u ∈ Rm : Lf h(x) + LF h(x)θ̂ + Lg h(x)u
is sufficient to conclude the safety of (25). o
≥ −α(h(x)) + ∥LF h(x)Γ∥1 + d∥∇h(x)∥ ¯ 2 ,
Lemma 3. Let h : Rn → R be an RaBF for (25) on a
set C ⊂ Rn as in (26). Then, (25) is safe on C for each (32)
(x0 , θ̂ 0 , Γ0 ) ∈ C × Rp × Rp×p . assigning to each triple (x, θ̂, Γ) the set of all control values
satisfying the condition in (31). The following theorem
Proof. The time-derivative of h is bounded as: shows that when θ ∈ Z(θ̂(t), Γ(t)) for all t ∈ R≥0 , any
ḣ =Lf h(x) + LF h(x)θ + Lg h(x)k(x, θ̂, Γ) + ∇h(x) · d locally Lipschitz controller drawn from (32) renders the
closed-loop system safe.
¯
≥Lf h(x) + LF h(x)θ + Lg h(x)k(x, θ̂, Γ) − d∥∇h(x)∥ 2
≥ − α(h(x)), Theorem 2. Let h : Rn → R be a RaCBF for (24) on
a set C ⊂ Rn as in (26). Let k : Rn × Rp × Rp×p →
where the final inequality follows from (27). Thus, since zero Rm be a locally Lipschitz controller satisfying k(x, θ̂, Γ) ∈
is a regular value of h and for all (x, θ̂, Γ, t), h(x) = 0 =⇒ K(x, θ̂, Γ) for all (x, θ̂, Γ) ∈ Rn × Rp × Rp×p , and let t 7→
ḣ(x, θ̂, Γ, t) ≥ 0, it follows from Nagumo’s Theorem [23, x(t), θ̂(t), Γ(t) be the trajectories of the resulting closed-
Ch. 4.2] that h(x(t)) ≥ 0 for all t ∈ R≥0 , which implies loop system (25) from a given initial condition (x0 , θ̂ 0 , Γ0 ) ∈
x(t) ∈ C for all t ∈ R≥0 , as desired. C × Rp × Rp×p . Then, if θ ∈ Z(θ̂(t), Γ(t)) for all t ∈ R≥0 ,
The previous lemma demonstrates that the satisfaction of the closed-loop system is safe on C.
inequality (27) is sufficient to conclude the safety of the Proof. Taking the derivative of h along the closed-
closed-loop system (25); however, directly verifying such an loop system trajectory and lower-bounding (omitting time-
inequality presents challenges since it depends on θ, which dependency for ease of presentation) yields:
is unknown. In what follows, we illustrate how the results on
zonotopes from the previous section can be used to construct ḣ =Lf h(x) + LF h(x)θ + Lg h(x)k(x, θ̂, Γ) + ∇h(x) · d
verifiable conditions for safety. To this end, note that if ¯
≥Lf h(x) + LF h(x)θ + Lg h(x)k(x, θ̂, Γ) − d∥∇h(x)∥ 2
Z(c, G) is a zonotope, then each z ∈ Z(c, G) satisfies the
≥Lf h(x) + LF h(x)θ̂ + Lg h(x)k(x, θ̂, Γ)
element-wise inequality [24]:
q q
¯
− ∥LF h(x)Γ∥1 − d∥∇h(x)∥ 2
X X
c− |Gi | ≤ z ≤ c + |Gi |, (28) ≥ − α(h(x)),
i=1 i=1
where the second inequality follows from Lemma 4 since
where |Gi | denotes element-wise absolute value. The above θ ∈ Z(θ̂(t), Γ(t)) for all t ≥ 0, and the third from (32).
property is now used to establish bounds on LF h(x)θ. The remainder of the proof follows that of Lemma 3.
Lemma 4. Suppose that θ ∈ Z(θ̂, Γ) ⊂ Rp . Then: Given a RaCBF, controllers enforcing safety of the closed-
n loop system can be synthesized via the quadratic program:
LF h(x)θ ≥ LF h(x)θ̂ − ∥LF h(x)Γ∥1 , ∀x ∈ R . (29)
Proof. Note than since {θ} = Z(θ, 0) ⊆ Z(θ̂, Γ), we have: min 1
2 ∥u − kd (x, θ̂, Γ))∥22
u∈Rm
Z(LF h(x)θ, 0) ⊆ Z(LF h(x)θ̂, LF h(x)Γ) ⊆ R. s.t. Lf h(x) + LF h(x)θ̂ + Lg h(x)u (33)
It then follows from (28) that: ¯
≥ −α(h(x)) + ∥LF h(x)Γ∥1 + d∥∇h(x)∥2,

z ≥LF h(x)θ̂ − ∥LF h(x)Γ∥1 , (30) where kd : Rn × Rp × Rp×p → Rm is a nominal feedback

controller. Note that safety of the closed-loop system under
for each z ∈ Z(LF h(x)θ̂, LF h(x)Γ). Since:
such a controller is contingent on the assumption that θ ∈
LF h(x)θ ∈ Z(LF h(x)θ̂, LF h(x)Γ), Z(θ̂(t), Γ(t)) for all t ∈ R≥0 , which holds if the update
laws in (10) and (11) satisfy the conditions of Theorem 1.
(30) implies (29), as desired.
We now have all the tools in place to introduce a con- Example 3. Consider an inverted pendulum with dynamics:
        
trolled version of a RaBF, which facilitates the synthesis of ẋ1 x2 0 0 θ1 0
= + + u + d(t)
controllers enforcing safety of the closed-loop system (25). ẋ2 0 sin(x1 ) −x2 θ2 1
| {z } | {z } | {z } |{z} |{z}
Definition 5. A continuously differentiable function h : ẋ f (x) F(x) θ g(x)
Rn → R defining a set C ⊂ R as in (26) is said to be a with state x = (x1 , x2 ) representing angular position and
robust adaptive control barrier function (RaCBF) for (24) velocity, input u ∈ R representing torque applied to the
on C if for all (x, θ̂, Γ) ∈ Rn × Rp × Rp×p : base of the pendulum, and θ = (θ1 , θ2 ) ∈ R2 representing
n o
sup Lf h(x) + LF h(x)θ̂ + Lg h(x)u uncertain parameters associated with gravitational effects and
u∈Rm (31) damping effects, respectively. The objective is to design a
¯
≥ −α(h(x)) + ∥LF h(x)Γ∥1 + d∥∇h(x)∥ 2 . controller such that the orientation of pendulum x1 satisfies
 1 <latexit sha1_base64="vyAxBR/pVnkJUZQ4G14S/xadrmY=">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</latexit>
set representations for uncertainty quantification, and con-
<latexit sha1_base64="2KyCcg4vbcq5m7UJV7xZZBJdqyY=">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</latexit>

h0 (x) = 0
<latexit sha1_base64="9z7loGSj6YP2bJv6VS6olpcfYQ4=">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</latexit>

= 0.1 (b) sidering stochastic disturbances.

0.8 4
x1 (t)

k✓(t)k
<latexit sha1_base64="24NMoWUFpA81/aXJCZesWPd5FWI=">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</latexit>

d(t) = 0
0.6 R EFERENCES
0.4 2

˜
<latexit sha1_base64="udP3qRvmg40crnTT2sFQRXKjiHo=">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</latexit>

= 1.5 [1] P. A. Ioannou and J. Sun, Robust Adaptive Control. Dover, 2012.
<latexit sha1_base64="mtKDWQlZU5Zywx4j8Y04fjFMR5Y=">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</latexit>

0.2
<latexit sha1_base64="XNowi0dpM8QdLf+3IDTbqeHSKgY=">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</latexit>

(a) [2] M. Krstić, I. Kanellakopoulos, and P. Kokotović, Nonlinear and

<latexit sha1_base64="2EKoQspXEabFep74H2PqwE55XX4=">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</latexit>
0 0 adaptive control design. John Wiley & Sons, 1995.
0 2 4 6 8 10 t <latexit sha1_base64="/EXVznWUhYIzIoiPeDG9UtuTRp0=">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</latexit>

0 2 4 6 8 10 [3] M. H. Cohen and C. Belta, Adaptive and Learning-based Control of

Safety-Critical Systems. Springer Nature, 2023.
Fig. 4. Trace of the pendulum’s position (a) and resulting parameter [4] A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath,
estimation error (b) for the simulations from Example 3. and P. Tabuada, “Control barrier functions: theory and applications,”
in Proc. Eur. Control Conf., pp. 3420–3431, 2019.
x1 ∈ [− π4 , π4 ], leading to the safety constraint h0 (x) = [5] A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier
( π4 )2 − x21 ≥ 0, which is used to construct a CBF using function based quadratic programs for safety critical systems,” IEEE
barrier backstepping [25]. As the dynamics are subject to Trans. Autom. Control, vol. 62, no. 8, pp. 3861–3876, 2017.
[6] A. J. Taylor and A. D. Ames, “Adaptive safety with control barrier
uncertain parameters and additive disturbances, we leverage functions,” in Proc. Amer. Control Conf., pp. 1399–1405, 2020.
the RaCBF approach developed in this section to design [7] B. T. Lopez, J. J. Slotine, and J. P. How, “Robust adaptive control
such a controller. To estimate the parameters online we barrier functions: An adaptive and data-driven approach to safety,”
IEEE Contr. Syst. Lett., vol. 5, no. 3, pp. 1031–1036, 2021.
leverage the approach from Sec. III using a history stack [8] A. Isaly, O. S. Patil, R. G. Sanfelice, and W. E. Dixon, “Adaptive safety
with M = 20 entries, where7 (1) is sampled every 0.1 with multiple barrier functions using integral concurrent learning,” in
seconds until the stack is full. We take our initial parameter Proc. Amer. Control Conf., pp. 3719 – 3724, 2021.
[9] M. H. Cohen and C. Belta, “High order robust adaptive control barrier
estimate as θ̂ 0 = (5, 2), reflecting the assumption that weaker functions and exponentially stabilizing adaptive control lyapunov
gravitational effects and more damping are present, and functions,” in Proc. Amer. Control Conf., pp. 2233–2238, 2022.
choose P = diag(0.1, 0.4) to reflect our confidence in such [10] M. H. Cohen and C. Belta, “Modular adaptive safety-critical control,”
in Proc. Amer. Control Conf., pp. 2969–2974, 2023.
an estimate. With such initial conditions, we take β = 0.1 [11] B. T. Lopez and J. J. Slotine, “Unmatched control barrier functions:
and simulate the closed-loop adaptive system (25) under Certainty equivalence adaptive safety,” in Proc. Amer. Control Conf.,
the influence of the controller in (33), where kd (x) = 0. pp. 3662–3668, 2023.
[12] M. Black, E. Arabi, and D. Panagou, “A fixed-time stable adaptation
The system is simulated from such initial conditions for law for safety-critical control under parametric uncertainty,” in Proc.
two different scenarios: i) without disturbances d(t) = 0; Eur. Control Conf., pp. 1328–1333, 2021.
ii) with disturbance d(t) = [0.0, −0.4 sin(t)]⊤ but where [13] M. Black and D. Panagou, “Safe control design for unknown nonlinear
systems with koopman-based fixed-time identification,” arXiv preprint
the disturbance is not accounted for in the controller (i.e, arXiv:2212.00624, 2022.
d¯ = 0 in (33)) or update law (i.e., via P and β), the [14] M. Tanaskovic, L. Fagiano, R. Smith, and M. Morari, “Adaptive
results of which are shown in Fig. 4 by the blue and red receding horizon control for constrained mimo systems,” Automatica,
vol. 50, pp. 3019–3029, 2014.
curves, respectively. In the disturbance-free case, safety is [15] V. Adetola, D. DeHaan, and M. Guay, “Adaptive model predictive
maintained and the parameter estimation error converges control for constrained nonlinear systems,” Syst. Control Lett., vol. 58,
to zero; when disturbances are present and not accounted no. 5, pp. 320–326, 2014.
[16] M. Lorenzen, F. Allgöwer, and M. Cannon, “Adaptive model predic-
for safety is violated. To enforce safety in the presence of tive control with robust constraint satisfaction,” IFAC-PapersOnLine,
disturbances, we rerun the system with β = 1.5 and d¯ = 0.4, vol. 50, no. 1, pp. 3313–3318, 2017.
the results of which are shown by the purple curves in Fig. 4, [17] G. Chowdhary, Concurrent learning for convergence in adaptive
control without persistency of excitation. PhD thesis, Georgia Institute
leading to a safe closed-loop system, although in a somewhat of Technology, Atlanta, GA, 2010.
conservative8 fashion. [18] G. Chowdhary and E. Johnson, “A singular value maximizing data
recording algorithm for concurrent learning,” in Proc. Amer. Control
Conf., pp. 3547–3552, 2011.
[19] A. Girard, “Reachability of uncertain linear systems using zonotopes,”
VI. D ISCUSSION AND C ONCLUSIONS in Proc. Int. Conf. Hybrid Syst. Comp. Control, pp. 291–305, 2005.
[20] A. Parikh, R. Kamalapurkar, and W. E. Dixon, “Integral concurrent
In this paper, we presented a framework for online parame- learning: Adaptive control with parameter convergence using finite
ter estimation and uncertainty quantification that unites tools excitation,” Int. J. Adapt. Control Signal Process., vol. 33, no. 12,
from set membership identification and concurrent learning pp. 1775–1787, 2019.
[21] M. Althoff, G. Frehse, and A. Girard, “Set propagation techniques for
adaptive control. Such developments were motivated by their reachability analysis,” Annu. Rev. Control Robot. Auton. Syst., vol. 4,
eventual integration with control barrier functions, which pp. 369–395, 2021.
ultimately allowed for synthesizing controllers enforcing [22] S. Sadraddini and R. Tedrake, “Linear encodings for polytope con-
tainment problems,” in Proc. Conf. Decis. Control, pp. 4367–4372,
safety of nonlinear systems with parametric uncertainties 2019.
and additive disturbances. We believe these results may [23] R. Abraham, J. E. Marsden, and T. Ratiu, Manifolds, tensor analysis,
serve as the starting point for various extensions, such as and applications. Springer, second ed., 1988.
[24] K. Ghasemi, S. Sadraddini, and C. Belta, “Decentralized signal tem-
integration with model predictive control, leveraging other poral logic control for perturbed interconnected systems via assume-
guarantee contract optimization,” in Proc. Conf. Decis. Control, 2022.
7 Here, we simply take y = ẋ − f (x) − g(x)u and ϕ = F(x), implying [25] A. J. Taylor, P. Ong, T. G. Molnar, and A. D. Ames, “Safe backstep-
ping with control barrier functions,” in Proc. Conf. Decis. Control,
w = d, where d may also capture noise present in measurements/estimates
pp. 5775–5782, 2022.
of ẋ.
8 This conservatism can be mitigated using the polytopic approach from
Remark 1 at the cost of additional computations.