IEEE TRANSACTIONS ON ROBOTICS, VOL.

40, 2024 4615

Adaptive Complexity Model Predictive Control

Joseph Norby , Member, IEEE, Ardalan Tajbakhsh, Student Member, IEEE,
Yanhao Yang , Graduate Student Member, IEEE, and Aaron M. Johnson , Senior Member, IEEE

Abstract—This work introduces a formulation of model pre-

dictive control (MPC), which adaptively reasons about the com-
plexity of the model while maintaining feasibility and stability
guarantees. Existing approaches often handle computational com-
plexity by shortening prediction horizons or simplifying models,
both of which can result in instability. Inspired by related ap-
proaches in behavioral economics, motion planning, and biome-
chanics, our method solves MPC problems with a simple model
for dynamics and constraints over regions of the horizon where
such a model is feasible and a complex model where it is not. The
approach leverages an interleaving of planning and execution to
iteratively identify these regions, which can be safely simplified
if they satisfy an exact template/anchor relationship. We show
Fig. 1. ACMPC selectively simplifies the model to promote efficiency without
that this method does not compromise the stability and feasibility sacrificing stability. For example, during a legged leaping task joint information
properties of the system, and measures performance in simulation may be required during takeoff and landing but can be omitted elsewhere without
experiments on a quadrupedal robot executing agile behaviors affecting the behavior.
over terrains of interest. We find that this adaptive method en-
ables more agile motion (55% increase in top speed) and expands
the range of executable tasks compared with fixed-complexity
implementations.
Index Terms—Dynamics, legged robots, optimization and hierarchy is vulnerable to failures caused by omitting portions
optimal control, underactuated robots. of the underlying model, raising a fundamental question of how
to balance model fidelity and computational efficiency.
Inspiration for answering this question can be drawn from
I. INTRODUCTION other scientific fields, in particular behavioral economics and
S DEMAND for robotic systems increases in industries, neuromechanics. The famous “Thinking: Fast and Slow” frame-
A such as environmental monitoring, industrial inspection,
disaster recovery, and material handling [1], [2], [3], so too has
work theorizes that human cognitive function can be described
with two systems, which, respectively, handle rapid, simple pro-
the need for motion planning and control algorithms that effi- cessing and slow, deliberative reasoning such that “[the complex,
ciently handle the complexity of their dynamics and constraints. slow system] is activated when an event is detected that violates
Legged systems in particular are well suited for these applica- the model of the world that [the simple, fast system] maintains”
tions due to their ability to traverse unstructured terrains with [4]. Extending this concept to the field of motion planning yields
behaviors, such as that shown in Fig. 1, yet they are so far largely meta-planning methods, which change their structure to leverage
restricted to conservative behaviors due to this complexity. A simple, fast models where possible and complex, slow ones
common approach to overcome these challenges is to break where the simple model is inaccurate [5], [6]. However, it is not
up the problem into a hierarchy of subproblems, which rea- well understood under what exact conditions a given dynamical
son over progressively shorter horizons with increasing model system may leverage a simple model without sacrificing stability
complexity. This hierarchy improves computational efficiency, and feasibility, nor is it clear when a more complex model should
which can be used to detect obstacles further away, react more be used without making the computational overhead intractable.
quickly to disturbances, or reduce energy costs. However, this To investigate these questions, we employ the templates
and anchors approach for analyzing model hierarchies [7].
Manuscript received 28 June 2023; revised 12 January 2024; accepted 21 April This framework describes relations between dynamical systems
2024. Date of publication 6 June 2024; date of current version 11 October 2024. where complex behaviors (anchors) can be captured by simpler
This paper was recommended for publication by Associate Editor J. Carpentier models (templates), and its connections to legged locomotion are
and Editor K. Mombaur upon evaluation of the reviewers’ comments. This
work was supported in part by the Chevron CTC, and in part by the National
well-studied in both animals and robots. From these observations
Science Foundation under Grant DGE-1745016 and Grant ECCS-1924723. we draw three following hypotheses.
(Corresponding author: Joseph Norby.) 1) From the templates and anchors relationship we can derive
The authors are with the Department of Mechanical Engineering, Carnegie sufficient conditions, which identify regions of a behavior
Mellon University, Pittsburgh, PA 15213 USA (e-mail: jcnorby@[Link];
amj1@[Link]). where complex dynamics and constraints can be safely
This article has supplementary downloadable material available at simplified without compromising stability and feasibility
[Link] provided by the authors. (i.e., regions where the robot is following the template
Digital Object Identifier 10.1109/TRO.2024.3410408
model).

1941-0468 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4616 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

2) Legged systems often operate in environments where these In general, the most common approach used to address the
conditions are satisfied during the majority of behaviors planning and control challenges of legged locomotion has been
and, therefore, stability and feasibility guarantees can through leveraging some kind of model reduction, wherein
often be retained even with simplified models. the problem is solved with a reduced subset of the state and
3) During behaviors in which these conditions are not dynamics. The solution of this simpler problem is then passed to
met, a controller which leverages adaptive complexity another system with a more complex model, which fills in the ad-
online can improve performance over fixed-complexity ditional details. The hierarchical nature of this approach enables
formulations by enabling more efficient motion planning optimization of each algorithm based on salient features of the
while retaining stability guarantees. problem, such as the timescales of the dynamics or rates of sensor
We evaluate these hypotheses by constructing a formulation information. Examples of this approach include efficient global
of model predictive control (MPC) that actively adapts the planners focused on exploration [14], [15], [16], [17], to local
model complexity to the task. This is achieved by iteratively planners that plan contact phases over a few gait cycles [18], to
identifying regions of the horizon where the behavior can be whole-body controllers with full-order representations over very
safely expressed with a simpler model. This formulation— small horizons [19], [20], [21]. While these hierarchies have
which we call adaptive complexity model predictive control been shown to be capable of producing dynamic locomotion,
(ACMPC)—is most efficient when the two models satisfy a they face a fundamental problem in that dynamics or constraints
relationship known as “exact anchoring” [8], and in the worst in the omitted space can render the desired motion suboptimal or
case converges to the standard all-complex MPC formulation. even infeasible. Conservative assumptions in the simple model
We evaluate Hypothesis 1 by showing that this algorithm pro- may fail to produce solutions in difficult environments, and
vides formal stability and feasibility properties with respect optimistic assumptions may lead to infeasible behaviors that
to the complex system. We evaluate Hypotheses 2 and 3 by the more myopic complex system may not recognize until too
applying this algorithm to a quadrupedal system and conduct- late.
ing simulation experiments on common environments, which Several methods have been explored to resolve the interface
legged platforms may encounter. These results show that the between these layers. One straightforward approach is to use the
majority of the behaviors in these environments admit feasible complex model to assess the simple solution, either by providing
simplifications and that the resulting improvement in computa- a boolean feasible/infeasible classification, computing some
tional efficiency enables a 55% increase in top speed compared value function, or indicating new search directions [22], [23],
with a system without these reductions. We also show that [24], [25]. This can be efficient since checking a solution in a
retaining knowledge of the dynamics and constraints in the higher order space is easier than searching for one, although it
complex system expands the range of executable tasks compared does not allow the simple model to directly reason about the
with a system, which uniformly applies these reductions. In dynamics or constraints in the complex space.
the particular case of legged leaping, this approach enables Another promising way to resolve this is to employ an
receding horizon execution of a body-length leaping behavior adaptive planning framework to reason over different models
while considering joint constraints, which expands the leap- based on the task and constraints. One flavor of this approach
ing ability over prior methods which do not consider these plans over a mixture of models of varying degrees of fidelity
constraints. with some predefined rules that guide when to use each and
The rest of this article is organized as follows. Section II how to transition between them [17], [26], [27]. Other ap-
covers related work in greater detail, and Section III formulates proaches leverage an adaptive composition of models with dif-
the problem and introduces notation. Section IV provides an ferent safety bounds to trade between performance and robust-
overview of the algorithm, whose formal properties are explored ness or expand the problem dimensionality as needed to find
in Section V. This algorithm is applied to legged locomotion collision-free paths. [5], [6], [28], [29], [30]. Our approach is
models in Section VI, and performance for these systems is more similar to the latter in terms of the underlying adaptation
quantified in Section VII. Finally, Section IX concludes this mechanism, but differs in that we derive the exact conditions
article and discusses limitations and future extensions enabled under which transitioning between models of varying fidelity
by this work. can be done without sacrificing stability and feasibility. Fur-
thermore, we demonstrate how such a mechanism can be ap-
plied to a receding horizon framework for online planning and
II. RELATED WORK
control.
Dynamic motion planning and control for systems with inter- Similar planning and control problems can also be solved
mittent contact is inherently difficult. However, enabling agile online using receding horizon methods. In particular, MPC is an
autonomy for such systems is critical for real-world applications iterative receding-horizon optimization framework that has been
that require the robot to touch the world. In particular, legged commonly used to solve constrained optimal control problems
robots have significant potential for real-world deployment. (OCPs) [31]. In the context of dynamic legged locomotion, MPC
But, they often suffer from difficulties arising from hybrid often computes feasible body and/or joint trajectories in order
dynamics, high state dimensionality, and nonconvex constraints to track a higher level reference plan while respecting dynamic,
on their kinematics and dynamics. Such problems render even state, and control constraints. Works such as [32], [33], and
basic motion planning problems polynomial space (PSPACE)- [34] compute the desired ground reaction forces using a single
complete [9], and as a result existing algorithms to solve them rigid body model, which are realized as joint torques using a
globally for dynamic legged systems cannot operate in real- whole-body controller. Although computationally efficient, this
time [10], [11], [12], [13]. approach typically uses a single simplified model with some

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4617

simple heuristics to approximate the motion of the legs. In TABLE 1

KEY SYMBOLS USED THROUGHOUT THIS ARTICLE WITH SECTION OR
comparison, this work blends a single rigid body dynamics EQUATION NUMBER OF THE INTRODUCTION MARKED
model with a full kinematics model, allowing us to more exactly
capture leg motion via joint limits, collision constraints, and
motor models. More generally, our approach is not tied to a
particular choice of models, but rather is a paradigm for selecting
between related models to improve performance.
While simplified model MPC approaches assume a represen-
tative reduced-order model (template) exists that fairly approx-
imates the full-order model (anchor), they often do not exam-
ine the validity of this approximation. Some approaches have
studied safe controller synthesis for the template model while
ensuring constraint satisfaction of the anchor model through
bounding the differences of the two models using reachability
analysis [35], approximate simulation properties [36], or learn-
ing any unmodeled differences [37]. Another approach uses pre-
defined ratios to mix the complex model for the immediate future
with the simple model for longer horizon planning, which results
in more robust locomotion compared with fixed-complexity
formulations [38]. Our approach takes inspiration from these in
that it defines the exact conditions under which the higher order
model can be simplified without violation of formal guarantees,
but critically differs in that we adaptively mix models of varying
fidelity within any planning horizon based on local feasibility.
This allows us to leverage the fidelity of the complex model
when necessary while taking advantage of the computation
speed enabled by the simple model for planning longer horizon
motions.
Many of these hierarchical planning and control approaches
have been shown to effectively perform agile and dynamic mo-
tions in simulation and hardware. However, they often struggle
in generating and executing motions that require the robot to
operate at its kinematic and dynamic limits. These motions are
critical in enabling behaviors, such as stepping and leaping over
gaps, stairs, and nontraversable obstacles, which are essential
in navigating unstructured terrains. Previous approaches lack
longer horizon planning, or do not consider joint kinematics or
constraints [39], [40], [41], [42], [43]. Our approach allows the
robot to plan for these agile behaviors in a receding-horizon
manner while reasoning about constraints over a significantly
long horizon, which is shown to expand leaping capabilities in
simulation.
Recently, some MPC approaches have been extended to gen-
erate dynamically feasible whole-body motion plans in real-
time [44], [45], [46], [47], [48], [49], [50]. Our approach is
not meant to be a replacement for this class of solutions, but
aims to complement them by providing systematic model reduc-
tion that reduces the computational load without compromising
feasibility. Prior work applying mixed-complexity schedules
to the structure-exploiting differential dynamic programming
(DDP)-based methods common in these works suggests that
the concepts are compatible [38]. This reduction in computa-
tional load can be beneficial in two ways. First, it can allow
other modules (e.g., perception, navigation, autonomy, etc.)
III. PRELIMINARIES
to use more resources as needed without exceeding the total
computational requirements of the system and lowers the total To clarify the operation of ACMPC and its properties,
power draw of onboard computation. Second, it could enable we define a formulation for MPC and the closed-loop sys-
MPC to deploy longer prediction horizons and, therefore, better tem it yields following [51] and using notation described in
reason about future constraints given a fixed computational Table I. Consider a nonlinear, discrete-time, dynamical system,
budget. which evolves on state manifold X under controls U and with

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4618 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

dynamics f functions of state and is strictly decreasing in time (for asymp-

totic stability) or bounded in magnitude by the control input
xk+1 = f (xk , uk ) (1) (for input-to-state stability). We borrow the standard definitions
where xk ∈ X , uk ∈ U are the current state and control at time of strictly increasing functions K and K∞ as well as Lyapunov
k ∈ Z, and xk+1 ∈ X is the successor state. Let Xf ⊆ X be the functions and asymptotic stability from [52].
set of all feasible states and Uf ⊆ U the set of feasible controls. For the closed-loop system in (4) to yield provable stability,
For the sake of notational simplicity, let zk := (xk , uk ) define a assume the following properties on VI (·), VT (·), and XT :
state-control pair such that xk+1 = f (zk ). Let the space of all Assumption 2: (a) Intermediate and terminal cost
such pairs be Z = X × U and the space of all feasible pairs be are upper and lower bounded by K∞ functions,
Zf = X f × U f . i.e., ∃αL,I , αU,I , αL,T , and αU,T ∈ K∞ such that the following
Let the set XN denote the basin of attraction of the controller holds.
parameterized by N . We list a few standard assumptions on the αL,I (|x|) ≤ VI (x, u) ≤ αU,I (|x|) ∀x ∈ Xf , u ∈ Uf
system in (1) and the set XN :
Assumption 1: (a) f (0, 0) = 0 (the origin is an equilibrium αL,T (|x|) ≤ VT (x) ≤ αU,T (|x|) ∀x ∈ XT .
point); (b) A solution to (2) exists for all xk ∈ XN .
(b) ∃u ∈ Uf | f (x, u) ∈ XN ∀x ∈ XN (XN is control positive (c) The functions VI (x, u), VT (x), and f are all twice differ-
invariant); entiable with respect to x and u;
(c) Xf and Uf are compact and contain the origin in their (d) ∃αVT ∈ K∞ , uT (x) : XT → Uf such that
interiors.
To formalize MPC, first define a predicted control tra- f (x, uT (x)) ∈ XT
jectory with horizon N and starting at time k as uk = VT (f (x, uT (x))) − VT (x) ≤ −αVT (|x|)
[u0|k , u1|k , . . . , ui|k , . . . , uN −1|k ] where i denotes the in-
dex within the horizon. In many cases the subscript ∀x ∈ XT (terminal control law decreases cost).
k can be inferred from context, allowing the sim- (e) The control law defined in (3) satisfies Assumption 1(a).
pler form u = [u0 , u1 , . . . , ui , . . . , uN −1 ]. Likewise let x = It is a known result that under these conditions, the origin of
[x0 , x1 , . . . , xi , . . . , xN ] denote the predicted state trajectory, the system defined in (4) is asymptotically stable for all x0 ∈ XN
and their pairing as z = [z0 , z1 , . . . , zi , . . . , zN −1 ] such that [31], [52]. These results can be extended to show the stability of
zi = (xi , ui ). Let the spaces of all state, control, and paired other stationary points through a coordinate transformation or
trajectories be denoted as X, U, and Z, respectively. The OCP the stability of a time-varying reference by transforming to an
solved in the standard NMPC formulation with terminal cost equivalent time-varying system [51].
and region is thus OCPN (xk ) : Xf → Z
IV. ADAPTIVE COMPLEXITY MPC
z∗k = OCPN (xk )
 N −1
 A. Algorithm Overview

= arg min VT (xN ) + VI (zi ) (2a) The core idea of adaptive complexity is to leverage models
zk of differing complexity to simplify the model in regions where
i=0
feasibility is assured and only increase complexity as needed to
s.t. xi+1 = f (zi ) ∀i = 0, . . . , N − 1 (2b)
guarantee feasibility and stability properties. Our approach is
zi ∈ Z f ∀i = 0, . . . , N − 1 (2c) to recursively define a simplicity set Sk at each time k whose
elements are times i in the prediction horizon where state-control
x0 = xk (2d) pairs can be simplified. This simplicity set is then used to define
xN ∈ XT (2e) a smaller OCP, which can be solved more efficiently and from
which a standard MPC control policy can be extracted. The OCP
where z∗k contains the optimal state and control trajectories x∗k solution is then used to update the next simplicity set, Sk+1 .
and u∗k , VI (·) and VT (·) are the intermediate and terminal costs, Conditions on this set Sk can be directly drawn from the
Zf is the set of feasible state-control pairs, and XT is a terminal literature on templates and anchors [7]—elements of Sk rep-
set. In most contexts the parameterization by N is implied, resent times corresponding to state-control pairs that follow a
and thus the problem will be denoted as simply OCP(xk ). The feasible, attracting, invariant submanifold within the complex
control law defined by NMPC is determined by solving (2) (“anchor”) space, i.e., follow the “template” dynamics. In other
at each time k and applying the first control, i.e., uk = u∗0|k . words, membership in Sk implies that the system remains on
This defines the state feedback policy umpc (xk ) and resulting the manifold after applying the complex dynamics and without
closed-loop system fmpc (xk ) as follows: violating constraints, as shown visually in Fig. 2. This approach
relaxes the typically restrictive requirement that template and
umpc (xk ) := u∗0|k (3) anchor relations are always satisfied for two dynamical systems,
fmpc (xk ) := f (xk , umpc (xk )). (4) and replaces it with a method to identify when they can be
satisfied at a given time. This allows the algorithm to account
Standard stability proofs for NMPC formulations typically rely for differences between the two systems, while improving per-
on showing that the closed-loop system admits a Lyapunov func- formance proportionally to the extent to which they align. Note
tion, which is upper and lower bounded by strictly increasing that while this work assumes one template per anchor, it could

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4619

feasible set of the simple system Zfs as follows:

xsk+1 = f s (zks ) (5)

zks ∈ Zfs . (6)

Let the basin of attraction of the complex system be denoted

XN c under horizon length N c .

Fig. 2. Elements in the horizon i are in the set Sk if they are feasible and C. Adaptive System Definition
stay on the manifold Z s (illustrated as a blue dashed 1-D curve). The adaptive
system allows zil to leave this submanifold while remaining in the manifold We seek an adaptive control law, which leverages the simple
Z c (the surrounding white 2-D space). Elements in the set Sk are denoted system when the system can feasibly remain on the manifold
in blue, elements not in this set are denoted in red. In this example, Sk = Zfs and the complex system when it cannot. Define another set
{1, 2, 3, 7, 8, 9}. of state and control variables xai , uai , and zia , which represent
an adaptive system used to solve the OCP. Note that this “sys-
tem” is only defined for the OCP, and thus is always indexed
be further extended to include multiple templates describing
by prediction horizon time i and not the dynamical system
different reductions of the anchor dynamics.
time k.
We will use the simplicity set Sk to assign these quantities at
B. Complex and Simple System Definitions each time i in the horizon to a particular system. More formally,
Let the original system defined in (1) be “complex,” which denote all combinations of horizon elements as S := {S | S ⊆
is clarified by the superscript (·)c . Let the “simple” system be {0, 1, . . . , i, . . . , N }}, and let Sk ∈ S. This simplicity set then
denoted with the superscript (·)s such that the state xs lies assigns a domain to each element of the horizon as follows:
on the manifold X s , where dim X s ≤ dim X c . These states  c
are related by the state reduction ψx : X c → X s defined as Z , i∈ / Sk
zia ∈ (7)
xs = ψx (xc ). Examples of these systems (which are further Z s , i ∈ Sk .
explored in Section VI) could be the full configuration space
of a legged robot or its centroidal momenta as X c , and some Often the adaptive states and controls at time i must be expressed
related system such as the single rigid body or inverted pendulum in the complex system whether or not i ∈ Sk . These instances
models as X s [53], where ψx is the projection—often via a can be handled by applying the lifting operator ψ † at time i if i ∈
selection matrix—from the complex space to the simple one. Sk . To simplify subsequent notation and clarify the difference
Note that the two systems could also share the same state space between a quantity in the complex system (i.e., zic ) and a quantity
but differ in feasible sets, for example, if one system has a more in the adaptive system, which may have been mapped to the
complex representation of constraints (motivating our choice of complex system (i.e., ψ † (zia )), we will use the shorthand zil ∈ Z c
the terms “complex/simple” over “full/reduced order”). to denote the mapped quantity, which is defined as follows:
Let each system have controls uc and us defined over man-  a
z , i∈ / Sk
ifolds U c and U s , which are related by the control reduction zil = i† a (8)
ψ (zi ), i ∈ Sk .
ψu : U c × X c → U s defined as us = ψu (uc , xc ). Define the
state-control pairs as z c = (xc , uc ) and z s = (xs , us ), which Next, we define the dynamics of the adaptive system, which
lie on manifolds Z c := X c × U c and Z s := X s × U s . This are used to solve the OCP as follows:
permits the definition of the reduction ψ : Z c → Z s defined
⎧ c a
as z s = ψ(z c ) = (ψx (xc ), ψu (uc , xc )). Examples of varying ⎪ f (zi ) i, i + 1 ∈/ Sk
⎨
fidelity control representations include actuator inputs for U c ψ x ◦ f c a
(z ) i ∈/ S k i + 1 ∈ Sk
,
xai+1 = f a (zia ) := i
(9)
and ground reaction forces, impulses, or impedances for U s , ⎪ f c
◦
⎩ s a i ψ † a
(z ) i ∈ S k, i + 1 ∈
/ Sk
where ψu again projects from the complex input to an equivalent f (zi ) i, i + 1 ∈ Sk
simple one.
While the many-to-one reduction ψ defines how the complex where xai+1 is the successor state in the adaptive system. The
and simple systems relate and is generally determined by the OCP for the adaptive system uses these dynamics to construct
choices of these two systems (e.g., a projection), more flexibility feasible motions over a prediction horizon N a ≥ N c .
is allowed in determining how to lift from the simple to the Denote a predicted adaptive control sequence over horizon
complex space. Define the mapping ψ † such that ψ ◦ ψ † = I, N a as ua = [ua0 , ua1 , . . . , uaN a −1 ], a predicted adaptive state
where I is the identity map. Let ψx† and ψu† give the outputs sequence as xa = [xa0 , xa1 , . . . , xaN a ], a predicted adaptive state-
of ψ † corresponding to state and control variables, respectively. control pair sequence as za = [z0a , z1a , . . . , zN a
a −1 ], and their

Examples of ψx† for legged systems include inverse kinematics respective spaces as Xa , Ua , and Za . Let the lifted form of
algorithms, which map a body state and reference foot tra- these trajectories be denoted as zl . We define the constraints
a
jectories to joint kinematics, whereas examples of ψu† include Zf,i in the adaptive system as follows:
controllers which map inputs from task space to actuator space.  c
The dynamics and constraints for the complex system have a Zf i ∈/ Sk
Zf,i := . (10)
already been defined in Section III. Define the dynamics f s and Zfs i ∈ Sk

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4620 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

D. Adaptive Complexity Optimal Control Problem

With these definitions in place we now state the adaptive com-
plexity optimal control problem (ACOCP)N a (xck , Sk ) : Xfc ×
S → Za , which is solved to determine the control input uck as
follows:
z∗a c
k = ACOCPN a (xk , Sk )
 N −1


= arg minzak VTc (xaN ) + VIa (zia ) (11a)
i=0

s.t. xai+1 = f a (zia ) ∀i = 0, . . . , N a − 1 (11b)

zia ∈ a
Zf,i a
∀i = 0, . . . , N − 1 (11c)
xa0 = xck (11d)
xaN a ∈ XTc (11e)
where the adaptive cost function is equal to the complex system
cost evaluated on the lifted state-control pair Fig. 3. ACMPC retains recursive feasibility and admissibility by updating the
 c a simplicity set and solution at time k + 1 with the corresponding terms from time
VI (zi ) i∈/ Sk k along with the new state and control determined from the terminal policy uT ,
VIa (zia ) = (12) which is applied with the terminal set XTc . Letters at the bottom indicate the
VIc (ψ † (zia )) i ∈ Sk . condition of (15) that element satisfies or violates. The complex manifold X c
In many contexts the ACOCP parameterization by N a and is the 2-D space, whereas X s is the embedded 1-D submanifold.
corresponding simplicity set Sk can be inferred, permitting
the shorthand ACOCP(xck ). Let XN a be the set of states for
which the solution to (11) exists. Let the optimal value func-
The condition (15a) requires that the first and last state be
tion found in (11) correspond to the control trajectory u∗a =
complex, which ensures the predicted trajectory matches the
[u∗a ∗a ∗a ∗l
0 , u1 , . . . , uN a −1 ] and corresponding lifted trajectory u . actual dynamics and that the system is able to reach the (possibly
The control law defined by ACMPC is determined by solving
complex) terminal state. The conditions (15b) require that the
(11) at each time k and applying the first control, i.e., uck = u∗l
0|k . state and control on the manifold at i are feasible in the complex
Define the state feedback policy ucacmpc (xck ) and resulting closed- space, (15d) requires the dynamics from the prior state and
c
loop system facmpc (xck ) as follows: control lead to the manifold, and (15c) requires that the dynamics
ucacmpc (xck ) := u∗l (13) applied to the current state yields a successor state on the mani-
0|k
fold, i.e., the complex space “exactly anchors” the simple space
c
facmpc (xck ) := f c (xck , ucacmpc (xck )). (14) (in the sense of [8, Appendix A]). These concepts are illustrated
in Fig. 3. To codify the admissibility of a simplicity set Sk , define
Note that each term in the ACOCP defined in (11) converges to its the set of all admissible sets as Sa ⊂ S. Membership in Sa can
complex counterpart if Sk = {}, so the behavior of the original be determined by computing the lifted prediction trajectory zl
MPC closed-loop system defined in (1) can always be retained. and checking the conditions in (15) for each i ∈ Sk . Also note
However, we seek to find the minimal complexity required to that removing any element from an admissible simplicity set
still ensure stability of the closed-loop system. does not invalidate the admissibility property.
In order to show the stability of adaptive complexity MPC,
E. Conditions on the Complexity Set we require that Sk be admissible for all k.
The set Sk clearly cannot be arbitrary in order to maintain Assumption 3: Sk ∈ Sa ∀k ≥ 0.
stability, as ignoring some uncontrolled component of the system The remaining challenge is ensuring the admissibility of Sk ,
dynamics could easily cause undesired behavior. To avoid this which is inductively handled by Algorithm 1 and for which
we define a notion of the admissibility of Sk , which is needed to recursive admissibility is proven in Section V-C, Lemma 7.
show that the state and control trajectories in the adaptive space There are multiple methods to ensure the initial S0 is admissible,
match their realizations in the complex space. such as initializing with S0 = {}, which is guaranteed to be
Definition 1: (Admissibility of Sk ) The simplicity set Sk admissible, or iteratively solving (11) and updating S0 until an
defined at time k is admissible if the following conditions hold admissible set is found. After an initial set is found, successor
for all i ∈ Sk : sets can be found by combining two simplicity sets, Ska and
S f . The set Ska adaptively identifies regions that can be safely
i = 0, N a (15a) simplified, whereas S f requires that certain elements always
†
ψ ◦ψ zil ∈ Zfc (15b) remain complex for stability and admissibility, in particular the
first and last element. After a solve, Ska is updated by checking
ψx† ◦ f s ◦ ψ zil = f c zil (15c) (15) to determine which states can be simplified. Its elements
a
are then shifted in time, i.e., Sk+1 ← {i | (i + 1 ∈ Ska ) ∨ (i =
ψx† ◦ ψ ◦ f c zi−1
l
= f c zi−1
l
. (15d) a
N )}, and combined with the fixed set to yield the successor

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4621

Algorithm 1: Adaptive Complexity Model Predictive Con- and show that the origin of the complex system is asymptotically
trol. stable by leveraging a cost function that penalizes deviations
Given xc0 , S0 , S f , N a from the origin. This “tracking MPC” formulation [51] requires
k←0 an additional assumption on how the heuristic lift ψ † and the cost
S0a ← S0 function VI interact, specifically that projecting to the simple
repeat manifold does not increase the cost.
Sk ← Ska ∩ S f Assumption 4: VIc (z c ) − VIc (ψ † ◦ ψ(z c )) ≥ 0.
z∗a ← ACOCPN a (xck , Sk ) (11) This can be easily obtained by assigning ψ † to map to the
z∗l ← ψ † (z∗a ) (8) origin for components in the null space of ψ, which by definition
uck ← u∗l (13) would have zero cost.
0|k
In the case of legged systems—such as those that will be
xck+1 ← f c (xck , uck ) (14) introduced in Section VI—or other systems that seek to track
Ska ← {i | i ∈ Ska ∧ zil satisfies (15)} reference trajectories rather than fixed points, the heuristic lift
a
Sk+1 ← {i | (i + 1 ∈ Ska ) ∨ (i = N a )} ψ † could map to the reference trajectory rather than the origin,
k ←k+1 resulting in a time-varying system. Also note that other formula-
until finished tions exist, which yield provable stability results for nontracking
cost functions—such as “economic MPC” [54]—and thus could
possibly yield provable stability under more relaxed assump-
a
simplicity set Sk+1 = Sk+1 ∩ S f . This approach is sufficient to tions on the form of ψ † . Theoretical analysis of these extensions
guarantee admissibility in the nominal case with a perfect model, will be left to future work to focus on the most straightforward
where the only new portion of the optimal trajectory is the last application of ACMPC.
element, which is always covered by S f . Robustness can be
improved at the expense of computational effort by expanding A. Optimal Control Problem Constraint Satisfaction
S f to include more elements to ensure Sk remains admissible
under disturbances. This is similar to the MPC formulation We begin by showing that admissibility of Sk results in a
in [38], but includes the adaptive term Ska to ensure feasibility lifted trajectory, which matches the evolution of the closed-loop
across the entire horizon. dynamics of the actual complex system, and therefore the con-
straints of the original OCP in (2) are satisfied by solutions of the
ACOCP in (11). Let the evolution of the closed-loop dynamics
F. Formal Definition of ACMPC Algorithm
starting at state xck under a given control trajectory uc for
With the control law and conditions for admissibility we can duration i be expressed by the function φf : Z × X c × U → X c
now summarize ACMPC as an iterative algorithm shown in defined as φf (i, xck , uc ) := f c (. . . f c (f c (xck , uc0 ), uc1 ), . . . , uci ).
Algorithm 1. This procedure combines the fixed and adaptive Proposition 1: Suppose Assumption 3 is satisfied. The pre-
simplicity sets, solves the ACOCP, extracts the lifted solution, dicted state at time i is equal to the solution to the complex
updates the adaptive simplicity set from this solution, then dynamical system under the lifted predicted controls, i.e., xli =
applies the first element of the control trajectory. Note that this φf (i, xc0 , ul ).
algorithm is no different from standard MPC formulations with Proof: We prove this by induction. For the base case i = 0,
the exception of the definition of the simplicity set and lifting of since 0 ∈ / Sk and xl0 = xc0 = φf (0, xc0 , ul ). For the induction
the resulting trajectory. step, we need to show that xli = φf (i, xc0 , ul ) implies xli+1 =
φf (i + 1, xc0 , ul ). We obtain φf (i + 1, xc0 , ul ) by applying the
V. THEORETICAL ANALYSIS closed-loop complex dynamics to φf (i, xc0 , ul ) with the control
determined by ul to both sides as follows:
This section describes the theoretical properties of ACMPC.
We first show that constraints of the original OCP in (2) are xli = φf i, xc0 , ul (16)
satisfied by solutions of the ACOCP in (11) under assumptions
on the admissibility of Sk (see Section V-A). We use this result to f c xli , uli = f c φf i, xc0 , ul , uli (17)
show recursive feasibility of the ACOCP, and thus asymptotic
stability of a point of interest of the closed-loop system (see f c zil = φf i + 1, xc0 , ul . (18)
Section V-B) under some assumptions on the form of ψ † . We
Thus, we need to show that xli+1 = f c (zil ) to verify that the
show that Algorithm 1 satisfies the assumption on admissibility
induction step holds. We proceed by cases based on inclusion in
of Sk (see Section V-C), and that the basin of attraction of the
Sk .
resulting system is no smaller than the original complex MPC
/ Sk , i + 1 ∈
Case 1. i ∈ / Sk : This case corresponds to a portion
system and possibly larger since the horizon length could be
of the trajectory entirely in the complex space. By the definition
expanded with the additional computational capabilities (Sec-
of the adaptive system dynamics in (9)
tion V-D).
Before delving into analysis, we must clarify the notion of xli+1 = xai+1 = f c (zia ) = f c (zil ). (19)
stability shown and the assumptions its proof requires. Recall
that stability for nonlinear dynamical systems is a local property Case 2. i ∈/ Sk , i + 1 ∈ Sk : This case corresponds to a portion
of an equilibrium point rather than a global property of the of the trajectory which decreases in complexity. By the definition
system. For simplicity, we will follow the proof presented in [31] of the adaptive system dynamics in (9) and the construction of

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4622 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

Sk in (15), (under Proposition 1). We now show that this control satisfies
the requirements for recursive feasibility.
xli+1 = ψx† (xai+1 ) = ψx† ◦ ψ ◦ f c (zia ) = f c (zia ) = f c (zil ) Proposition 4: Suppose, Assumptions 1–3 are satisfied. Let
(20) xck ∈ XN a and let xck+1 := facmpc
c
(xck ) denote the successor state
c a c
(under ACMPC) to xk . Then, ũ (xk ) defined in (23) is feasible
Case 3. i ∈ Sk , i + 1 ∈
/ Sk : This case corresponds to a portion c
for ACOCPN a (facmpc (xck )) and XN a is positively invariant (for
of the trajectory which increases in complexity. By the definition the system xk+1 = facmpc (xck )).
c c
of the adaptive system dynamics in (9) and the construction of Proof: This proof follows standard methods for demonstrat-
Sk in (15) ing recursive feasibility [31], see Appendix A for the complete
proof.
xli+1 = xai+1 = f c ◦ ψ † (zia ) = f c zil . (21) We must also show that the cost function decreases along any
solution of xck+1 = facmpc
c
(xck ) given these previous assumptions
Case 4. i ∈ Sk , i + 1 ∈ Sk : This case corresponds to a portion as this is necessary for the stability proof.
of the trajectory entirely in the simple space. By the definition Proposition 5: Suppose, Assumptions 1–4 are satisfied. Then
of the adaptive system dynamics in (9) and the construction of
Sk in (15) VN∗a facmpc
c
(xck ) − VN∗a (xck ) ≤ −VIc xck , ucacmpc (xck ) . (25)

xli+1 = ψx† (xak+1 ) = ψx† ◦ f s (zia ) = ψx† ◦ f s ◦ ψ(zil ) = f c zil . Proof: See Appendix A. 
(22) We can now prove asymptotic stability of the origin of the
closed-loop system using standard Lyapunov-based methods.
Thus, the induction step holds, completing the proof.  This supports Hypothesis 1 which states that adaptive complex-
We must also show that state and control trajectories, which ity MPC yields provable stability properties reliant on template
satisfy the adaptive state and control constraints Zfa also satisfy and anchor conditions.
the complex equivalent Zfc by nature of the admissibility of set Theorem 6: Suppose, Assumptions 1–3 are satisfied. Then,
Sk . there exists functions α1 , α2 , α3 ∈ K∞ , which upper and lower
Proposition 2: Suppose, Assumption 3 is satisfied. If zia ∈ bound the cost, i.e.,
Zf , then zil ∈ Zfc .
a
α1 (|xck |) ≥ VN∗aa (xck ) ≥ α2 (|xck |) (26a)
Proof: Proceed by cases based on inclusion in Sk .
/ Sk : In this case, zil = zia . By the definition of Zfa
Case 1. i ∈ VN∗aa (xck+1 ) − VN∗aa (xck ) ≤ −α3 (|xck |) (26b)
in (10), zi ∈ Zfc .
l
and thus the origin of the system
Case 2. i ∈ S: In this case, zil = ψ † (zia ). Since Sk ∈ Sa , zil ∈
c
Zf by (15b). xck+1 = facmpc
c
(xck ) (27)
Next, we show that satisfying the initial and terminal state
constraints in the adaptive system implies satisfaction of is asymptotically stable with a region of attraction XN a .
the same constraints in the complex space, which is trivial Proof: See Appendix A. 
since the initial and terminal states will always be expressed
in the complex space by the construction of Sk . C. Recursive Admissibility of Sk
Proposition 3: Suppose, Assumption 3 is satisfied. If xa0 = Theorem 6 shows that ACMPC is stable under Assumption 3,
xck then xl0 = xck , and if xaN a ∈ XTc , then xlN a ∈ XTc . which states that Sk is admissible. Next, we prove that this holds
Proof: Since i = 0 ∈ / Sk and xl0 = xa0 = xck . Since i = N a ∈ / for all time under Algorithm 1, again assuming no modeling
Sk and xN a = xN a ∈ XTc .
l a
 errors. Robustness considerations remain an intriguing area for
future investigation.
B. Adaptive Complexity Feasibility and Stability Lemma 7: Let xc0 ∈ XN a . Then, Sk ∈ Sa ∀k ≥ 0 under Al-
gorithm 1.
With these propositions in place we can now prove that the
Proof: We proceed by induction. The base case k = 0 is
OCP defined in (11) is recursively feasible for states in XN a
met by the assumption that S0 ∈ Sa . For the induction step we
and that this set is invariant in the complex system. This is done
must show that Sk ∈ Sa implies Sk+1 ∈ Sa . By Proposition 4,
by following the form of the proofs in [31], which constructs ∗l
zk+1 consists of each of the last N a − 1 elements of zk∗l , plus
a feasible solution (in the absence of modeling errors) to the
the new terminal state-control pair (xT , uT (xT )). Since under
OCP at the successor state by combining the current solution a
Algorithm 1 elements of Sk+1 are the time-shifted elements
with the terminal set feedback policy uT . This approach is a
of Sk , which satisfy the admissibility conditions (15), these
illustrated in Fig. 3. Define this control sequence as ũa and its a
conditions are satisfied for all i ∈ Sk+1 . Since by the definition
lifted counterpart ũl as follows:
of Sk+1 = Sk+1 ∩ S where i = N a ∈
a f
/ S f , the new terminal
ũa (xck ) = u∗a ∗a
1 , . . . , uN a −1 , uT (xT ) (23) state-action pair is always in the complex space, and hence (15)
are satisfied for all i ∈ Sk+1 and thus Sk+1 ∈ Sa . 
l
ũ (xck ) = ψu† (ũ a
(xck )) (24) Note that the terminal state-control pair may not meet the
reduction conditions in (15), meaning that it must remain in
where xT = φf (N a , xck , u∗l ) is the terminal state at time N a re- the complex space. This is handled by assuming the last finite
sulting from initial state xck and lifted control trajectory u∗l (xck ) element in the horizon is complex, and checking (15) after

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4623

solving the OCP to determine if the new index can be allowed

into Sk or must remain complex.

D. Basin of Attraction Comparison

We have shown that both the original MPC formulation for
the complex system in (4) and the adaptive complexity MPC
system in (14) are asymptotically stable about the origin with
basins of attraction XN c and XN a , respectively. We now show
that the size of their basins of attraction is dependent on the
horizon lengths N c and N a .
Lemma 8: If N c ≤ N a then XN c ⊆ XN a , and if N c < N a
then XN c ⊂ XN a .
Proof: Let xc0 ∈ XN c , thus OCPN c (xc0 ) has a solution u∗
satisfying the constraints in (2). Let S0 be the initial admissibility Fig. 4. Complex model includes body and foot states which together define
joint states, whereas the simple model only consists of body states. (a) Complex
set of this solution determined by evaluating the conditions in system. (b) Simple system.
(15). Construct an adaptive control trajectory ua as follows:
 ∗
ui i∈/ S0
uai = (28)
ψu (u∗i ) i ∈ S0 if this constraint were lifted. As such this approach may result
in a sacrifice in cost optimality in favor of a simpler problem.
for all i = 0, 1, . . . , N c − 1, which implies ul = u∗ . By Propo-
sitions 1–3 the constraints on the OCPs (2) and (11) are equiv- VI. APPLICATION TO LEGGED SYSTEMS
alent, and thus ua is a valid solution of (11), which implies that To demonstrate the validity of the proposed ACMPC in con-
xc0 ∈ XN a and thus XN c ⊆ XN a . trolling dynamical systems and to provide examples for the
To show that N c < N a → XN c ⊂ XN a , consider a point quantities defined above, we apply the approach for a legged
c
x0 ∈ / XN c and some feasible control uc0 ∈ Uf such that z0c := robot system. This section defines the complex and simple
(x0 , uc0 ) ∈ Zfc and xc1 = f c (z0c ) ∈ XN c . Let u∗1 be the control
c
models, which are used for implementing the algorithm, as well
sequence yielded by solving OCPN c (xc1 ). Let the control se- as the mappings between them. The complex model includes
quence ua0 = [uc0 , ua1 ], and let N a = N c + 1. Since the state states and constraints on both the body and feet of the robot
xc1 ∈ XN a , by Propositions 1–3 the state, control, and terminal (and thus also joint information through kinematics calculations)
constraints of (11) are satisfied for zia ˜∀i = 1, 2, . . . , N c , and whereas the simple model only considers body motion.
since z0c ∈ Zfc , the state and control constraints are satisfied for
i = 0. Thus, all the constraints of (11) are satisfied, and therefore
xc0 ∈ XN a . The property XN c ⊂ XN a follows.  A. Definition of Complex Legged System
The property that longer horizon lengths yield larger basins The complex system represents a model of the robot that
of attraction is a known result of MPC theory–as horizon includes body and foot states as shown in Fig. 4. More precisely,
lengths go to infinity, MPC converges to infinite-horizon op- we leverage a single rigid body dynamics model and a full kine-
timal control [31]. Horizon lengths are generally limited by matics model, similar to [13]. We choose to include foot states
computational effort, so if simplifying the problem allows for in the dynamical system rather than joint states for two reasons.
longer horizons under equal computational capabilities without First, we note that joint information is available in the OCP
violating system constraints, the resulting controller has a larger via slack variables constrained by forward kinematics, and that
basin of attraction. We note this theoretical result, although in operating in end effector space renders certain functions–in par-
the following experiments we leave horizon length fixed and ticular approximated limb dynamics and terrain clearance–linear
allow solve time variations rather than the reverse as this yields with respect to state variables and, therefore, more conducive
a more consistent representation of computational constraints. to online replanning. Second, this parameterization enables a
Also note that since membership in the basin of attraction implies quasimassless leg assumption, where limb dynamics can still be
that a solution to the OCP exists, Lemma 8 implies that adaptive approximated (in this case with a point-mass model) without
complexity maintains the completeness properties of the original affecting body dynamics, which is desirable to satisfy (15c) as
MPC formulation in the complex system [assuming a suitable much as possible. Note that joint dynamics could be accounted
algorithm which can solve the OCP in (11)]. for by using a centroidal dynamics model for the simple system
In addition, it should be noted that although feasible solutions and whole-body dynamics as the complex system. This approach
of the ACOCP are also feasible solutions of the original complex would allow the simple model to ignore joint dynamics when
OCP, they may not be optimal with respect to the original OCP feasible while still capturing the effects of limb inertia and
cost function. Reducing the system at a particular time is in kinematics on body dynamics, and only compute joint dynamics
essence constraining it to the simple manifold at that time. The when needed to satisfy actuation constraints. However, since the
conditions stated above require that doing so be feasible and massless leg assumption has been shown to be useful for systems
yield a decreasing cost, yet it is possible that doing so could such as the quadrupedal one studied here [53], [55], and we
yield a trajectory of higher cost across the whole trajectory than are primarily interested in behaviors where limb kinematics are

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4624 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

more relevant than their dynamics, we leave the extension to positions, velocities, and torques, respectively, where nj is the
other systems as future work. number of joints per leg. These slack variables are constrained
We define the states of the complex system xc ∈ R12+6n with
FK(q c , θ) = qfoot (33a)
n denoting the number of legs as follows:
 T
xc = qlin
T T
qang T
qfoot T
q̇lin ωT T
q̇foot
T
(29) J(q c , θ) q̇lin
T
, ω T , θ̇T = q̇foot (33b)

where qlin ∈ R3 defines the body position in the world frame, τ = −Jθ (q c , θ)T ubody (33c)
qang ∈ R3 defines the body orientation through a local param-
where FK(q c , θ) is the forward kinematics function of the sys-
eterization such as Euler angles, qfoot ∈ R3n defines the foot
tem, J(q c , θ) is the leg Jacobian which relates motion of the
positions in the world frame, and ω ∈ R3 is the angular velocity
body and joints to foot motion in the world frame, and Jθ are the
in the body frame (with ω  its skew-symmetric equivalent). Let
columns of this matrix corresponding to joint motion. Note that
the configuration be defined as q c := [qlin
T T
, qang T T
, qfoot ] .
c 6n
(33c) omits joint dynamics under the massless leg assumption.
Let the control inputs u ∈ R be defined as follows: As discussed at the beginning of this section, these additional
 
u variables and constraints allow the optimization to determine
uc = body (30) joint kinematics, which correspond to a given set of body and
ufoot
foot kinematics. This in turn enables satisfaction of kinematic
where ubody ∈ R3n are the desired ground reaction forces at each limits without requiring their direct inclusion in the dynamical
foot in the world frame coordinates, and ufoot ∈ R3n denote the system model.
forces which accelerate the feet during swing, but do not act on With these slack variables and their corresponding constraints
the robot body. Note that these forces on the physical system in place, we can now state the constraints in the complex system
correspond to the same actuators as any given leg is either as follows;
in stance or swing–this separation is primarily to distinguish
θmin ≤ θ ≤ θmax (34a)
between control authority available in the simple and complex
systems. θ̇min ≤ θ̇ ≤ θ̇max (34b)
Let ubody,j ∈ R3 and qfoot,j ∈ R3 be the ground reaction
forces and foot positions, respectively, for stance leg j. Using a τmin ≤ τ ≤ τmax (34c)
simple point-mass model for the feet and following the standard ubody,min ≤ ubody ≤ ubody,max (34d)
formulation for single rigid body dynamics models, e.g., as
in [56], define the continuous time dynamics of this system ubody ∈ FC (34e)
f c (xc , uc ) as follows: Dubody = 0 (34f)
⎡ ⎤
q̇lin D̄q̇foot = 0 (34g)
⎢ Jω (qang )ω ⎥    
⎢ q̇ ⎥
⎢ ⎥ θ̇ θ̇
f c (xc , uc ) = ⎢ 1 
foot
⎥ (31) −τmax 1+ ≤ τ ≤ τmax 1− (34h)
⎢ m j ubody,j − g ⎥ θ̇max θ̇max
⎣W (q , q , ω, u )⎦
lin foot body
ufoot h(qfoot ) ≥ 0 (34i)
where Jω (qang ) ∈ SO(3) is the linear mapping from angular where (·)min and (·)max represent variable bounds, (34e) enforces
velocity to the derivative of the orientation local parameteri- that the ground reaction force (GRF) at each foot lies within
zation, m is the body mass, g is the gravity vector, and the the linearized nonadhesive friction cone FC, (34f) and (34g)
shorthand function W (·) maps the state and control to the enforce a contact schedule with selection matrix D and its
angular acceleration as follows: bitwise complement D̄, (34h) enforces a linear motor model,
and (34i) enforces nonpenetration of the terrain via the ground
W (qlin , qfoot , ω, ubody ) clearance h(qfoot ). Note that GRF bounds are retained to ensure
 nonadhesion and discourage abuse of kinematic singularities.
−1 T
=M R(qang ) ((qfoot,j − qlin ) × ubody,j ) − ω
M ω
Together the constraints in (34) define the set Zfc .
j
Lastly, we must define the cost functions for the OCP. While
(32)
this is still an active area of research for legged systems, a
where M is the inertia matrix in the body frame and R(qang ) ∈ common approach is to extract a reference trajectory either from
SO(3) maps vectors from the body frame to the world frame. a higher level motion planner or by integrating forward a desired
The discrete time formulation of the dynamics in (31) can be body velocity, supplement it with task-space foot trajectories,
obtained with a suitable integration scheme such as forward then penalize any deviations [44], [47], [48], [57]. Let those
Euler. trajectories be described by x̄c , ūc . Under this approach, the
Next, we define the constraints of the complex system Z c . intermediate and terminal costs for the OCP of the complex
Kinematic constraints for legged systems are generally functions system can be defined as follows:
of joint limits rather than body or foot variables. As a result, we ecx,k = xck − x̄ck (35a)
add the joint information to the OCPs as slack variables and
use them to define constraints. Let θ, θ̇, τ ∈ Rn·nj be the joint ecu,k = uck − ūck (35b)

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4625

VIc (xck , uck , k) = ecx,k T QI ecx,k + ecu,k T Recu,k (35c) reference z̄c . If the reference is time-varying (as is the case for
legged systems), this results in time-varying lifts as follows:
VTc (xck , k) = ecx,k T QT ecx,k (35d)
⎡ ⎤
where ecx,k and ecu,k are, respectively, the state and input errors I3 0 0 0
from the time-varying reference, and where QI , QT , and R are ⎢0 I3 0 0 ⎥
⎢ ⎥
⎢0 0 0 0⎥ s
positive definite matrices. Note that while these matrices must be ψx† (xs , k) = ⎢ ⎥x
positive definite to yield provable stability, often assigning very ⎢0 0 I3 0 ⎥
⎣0 0 0 I3 ⎦
low costs to particular components will yield behavior more
consistent with online motion planners than tracking controllers 0 0 0 0
(leveraging available degrees of freedom to minimize the overall ⎡ ⎤
0 0 0 0 0 0
cost). ⎢0 0 0 0 0 0 ⎥
⎢ ⎥
⎢0 0 I3n 0 0 0 ⎥ c
+⎢ ⎥ x̄ (39a)
B. Definition of Simple-Legged System ⎢0 0 0 0 0 0 ⎥ k
⎣0 0 0 0 0 0 ⎦
The simple system represents the reduced-order model of the
0 0 0 0 0 I3n
robot that uses only the body states and ignores the foot and    
joint states, as shown in Fig. 4. This is a commonly used model I 0 0
ψu† (us , k) = 3n us + ūck . (39b)
reduction technique in the legged locomotion literature. The 0 0 I3n
state of the simple system is defined as xs ∈ R12 consisting
of only the body position, orientation, and velocities as follows: Note that using dynamically consistent reference trajectories
when defining this lift increases the likelihood that simplicity
T
xs = qlin
T T
qang T
q̇lin ωT . (36) set conditions will be met, since they require that states on the
simple manifold will remain there. Also since we will often want
The control inputs u ∈ R3n of the simple system consists of to compute equivalent joint data to accompany these null space
only the GRFs from the complex system, such that us = ubody . variables, these functions can be postprocessed with inverse
The dynamics in the simple system are the components of the kinematics and inverse dynamics to compute the values of θ,
complex dynamics corresponding to the simple system states θ̇, and τ that correspond to a particular value of zkc .
defined as f s (xs , us , k) as follows: We can now state the conditions for admissibility of this
⎡ ⎤ system, which require that the variables in the null space of
q̇lin ψ lie on the reference and are feasible.
⎢ Jω (qang )ω ⎥ Lemma 9: For a given state-control pair zil , which lies on
f s (xs , us , k) = ⎢⎣ m 1
 ⎥.
⎦ (37) a trajectory of the system defined in (31), a reduction at i ∈
j u body,j − g
W (qlin , q̄foot,k , ω, ubody ) [1, . . . , N − 1] is admissible (see Definition 1) if qfoot = q̄foot ,
q̇foot = q̄˙foot , ufoot = ūfoot , and zil satisfies the constraints in (34).
Note that in the simple system, q̄foot,k is a parameter (not a de- Proof: See Appendix B. 
cision variable), which is necessary for the angular acceleration
update, and renders the dynamics time-varying. The constraints D. ACMPC Algorithm Application
in the simple space are only constraint bounds on the input ubody ,
identical to (34d)–(34f). We can now restate the major components of the ACMPC
algorithm in the context of this example, specifically the ACOCP
and the simplicity set update.
C. Relations Between Complex and Simple-Legged Systems 1) Legged System ACOCP: The components needed to solve
With these systems defined, we can now relate the two with the ACOCP defined in (11) are the dynamics, constraints, costs,
the reductions ψx , ψu and define our heuristic lifts ψx† , ψu† . The and terminal conditions.
reductions select components to retain within the simple system The adaptive dynamics f a match the form of those specified
as follows: in (9), but leverage the equations defined above, specifically the
⎡ ⎤ complex dynamics f c in (31), the simple dynamics f s in (37),
I3 0 0 0 0 0 the reduction ψ in (38), and the heuristic lift ψ † in (39).
⎢ 0 I3 0 0 0 0⎥ c These equations together form the optimization constraint
ψx (xc ) = ⎣ x = xs
0 0 0 I3 0 0⎦
(38a)
(11b), which is added in a piecewise manner to each finite
0 0 0 0 I3 0 element in the ACOCP. To maintain admissibility properties,
we must also enforce that complex-to-simple system transitions
ψu = I3n 0 uc = us (38b)
always result in successor states that reside on the simple man-
where Im is an m × m identity matrix and 0 is a matrix of zeros ifold. This can be done by lifting the transition as follows:
of corresponding size. ψx† (xai+1 ) = f c (zia ) i∈
/ S k , i + 1 ∈ Sk . (40)
As discussed in Section V, a sensible choice for the heuristic
lift ψ † is a mapping, which preserves the states in the simple Practically speaking, this encodes the null space of ψ as opti-
system and fills in null space components with values from the mization parameters (i.e., q̄foot ) rather than decision variables

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4626 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

(i.e., qfoot ) when evaluating the dynamics constraint defined by only the simple and complex model dynamics and constraints,
(31), which effectively requires the control input to drive the and “Mixed” employs the complex model for the first one-
state exactly onto the simple manifold. quarter of the horizon and the simple model for remainder, sim-
Constraints are defined based on the simplicity set as de- ilar to [38]. The “Adaptive” configuration follows Algorithm 1
scribed in (10), where Zfc is the set of all state-control pairs with S f = {2, . . ., N a − 2} so that the first and last two knot
that satisfy (33) and (34), and Zfs is the set of pairs that satisfy points—and thus the first and last finite elements—are always
(34d)–(34f). These constraints are then applied piecewise as in complex. This slight restriction of the fixed simplicity set is
(11c). added to improve robustness to modeling errors as discussed
Note that this means complicated joint constraint functions in Section IV-E. It also ensures that new elements entering the
are only evaluated for finite elements in the complex system. horizon will be complex, which is needed to meet Assumption 3
Costs are likewise defined as described in (12) and applied in lieu of a hard-to-find terminal controller uT that meets the
piecewise as in (11a), with VIa computed with the cost function conditions in Assumption 2.
VIc defined in (35), and the heuristic lift ψ † in (39). In each experiment, the robot is provided a reference trajec-
The initial condition constraint in (11d) is applied to the first tory, which defines the required task over the given environment.
finite element, which is guaranteed to be in the complex space. We use three environments to evaluate the algorithm perfor-
There are many options for the terminal set constraint in (11e). mance in the presence of varying constraints. The “Acceleration”
In the absence of a terminal control law uT —which has no environment requires the robot to accelerate and decelerate over
known form for legged systems—common approaches include 7.5 body lengths (3 m) of flat terrain as quickly as possible to
constraining the final state to lie on the reference, or increasing measure the ability to stabilize the system during agile motions.
the terminal cost VTc to strongly penalize solutions which do not The “Step” environment consists of a one-half leg length (20 cm)
drive the system to the reference. step, which requires navigating state constraints such as joint
These equations fully specify all the components in (11), limits to traverse. The “Gap” environment consists of a body-
which can be evaluated in a piecewise manner by an off-the- length (40 cm) gap, which the robot must leap across, testing
shelf nonlinear program (NLP) solver. This work transcribes the controller’s ability to handle these kinematic constraints
the NLP using direct methods and solves with interior point in addition to input constraints, such as actuator limits and
optimizer (IPOPT) [58], but other works have solved similar friction.
mixed-complexity problems with indirect, DDP-based meth- Performance is quantified in each experiment by measuring
ods [38]. Note that the mappings ψ and ψ † are only needed to success rate, mean solve time, slow solve rate (% of solves
evaluate the dynamics constraint when transitioning onto or off that take longer than one timestep), and a task-specific metric.
the simple manifold or when evaluating the cost for the simple Success is defined as reaching the goal within a one-half body
system. length (20 cm) and zero velocity without failing to solve the NLP.
2) Legged System Simplicity Set Update: Once the ACOCP Data is recorded from the instant the first NLP is solved until
in (11) is solved, the next step in Algorithm 1 is to evaluate either the success or failure criteria are met, and only included
the admissibility conditions. Conditions (15c) and (15d) which in metrics if successful. In the Acceleration experiment, the
regulate transitions onto and off the manifold are automatically task-specific metric is average top speed achievable with a 100%
enforced via the heuristic values in the dynamics constraint as success rate over ten trials, which is determined by iteratively
described in (40), so the only condition that must be checked decreasing the reference duration until the robot was unable to
is the feasibility condition in (15b). This is done by first ob- maintain a 100% success rate. The task-specific metric for the
taining ψ † ◦ ψ(z∗l ), which uses the maps in (38) and (39) Step and Gap experiments is mean control error, which is the
to project the foot states and controls to the heuristic (along norm of the error between the sum of ground reaction forces
with joint states and controls through inverse kinematics and and the weight of the robot, averaged over the successful trials.
dynamics). These state-control pairs can then be checked for For the Step and Gap environments, the robot is initialized to a
feasibility in the complex system using (34). If any constraints random position within one-half body length in the transverse
are violated at a time i, then both i and i − 1 are removed plane from a nominal position. In each configuration, solve time
from the simplicity set. Removing both knot points ensures corresponds to the time spent in the NLP solver to isolate the
that a full finite element of complexity will be added to the effects on the underlying OCP, although the time required to
problem, which enables states to leave the simple manifold evaluate the conditions in (15) is negligible—our unoptimized
(see Fig. 2 for intuition behind this principle). Otherwise, all implementation takes around 1 ms in the worst case.
conditions in (15) are met and i can be added to the simplicity In each environment a fixed reference trajectory for the body
set, at which point the control is executed and the algorithm is provided from the global planner described in [17], and the ref-
repeats. erence contact sequence and foot trajectories are chosen online
before each MPC iteration with a Raibert-like heuristic [59] and
a threshold on traversability, as described in [57]. Note that these
VII. EXPERIMENTAL EVALUATION
references could have been provided as twist inputs integrated
This section presents experiments deploying adaptive com- forward in time from the current state—we chose to provide fixed
plexity MPC on a simulated quadrupedal robot to quantify references so they would remain invariant across all trials. We fix
its performance and benchmark against other formulations of the prediction horizon at two gait cycles (N c = N a = 24) with
MPC. In particular we compare against three other model a timestep of Δt = 0.03 s and measure the relative solve times as
configurations—“Simple” and “Complex,” respectively, employ discussed in Section V-D, although future work could implement

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4627

an adaptive horizon approach to keep solve time fixed. For each A baseline version of this experiment was run to remove the
solve, we use the state provided by the simulator as the initial effect of solve time on locomotion stability. In this experiment,
state. the simulation update rate was slowed down by a factor of
Once the reference and state information have been obtained, 20, and a maximum effective planning rate of 100 Hz was
we construct the NLP with the appropriate complexity structure enforced such that regardless of solve time, new solutions were
and solve it with IPOPT [58]. We configure IPOPT to enable published at that rate. With these modifications, the maximum
warm start initialization and provide it the primal and dual velocities were 3.2 m/s for Simple, 2.7 m/s for Complex, 2.9 m/s
variables from the prior solve (appropriately shifted) for rapid for Mixed, and 2.7 m/s for Adaptive. All configurations were
convergence. In the event that complexity must be unexpectedly within 15%, and constrained by the reach of the swing legs.
inserted in the middle of the horizon (discussed in more detail Notably, the Simple configuration still performed best because
in Section VIII), warm starting is disabled and nominal values it could violate these kinematic constraints in ways that the more
are used for the initial guess of the newly inserted null-space complex configurations could not. These effects are discussed
variables. Once the problem is solved, the MPC control output in greater depth in Section VIII.
is then mapped from ground reaction forces to joint torques
via the Jacobian-transpose method, and the resulting swing
foot trajectories are tracked with proportional-derivative (PD) B. Step Environment
control. See [57] for more details on the implementation of
A series of snapshots of the adaptive configuration navigating
the low-level controller. All experiments were performed using
the Step environment are shown in the top row of Fig. 6(a). The
Gazebo 9 with the ODE physics engine, and all processes were
key constraints, which must be resolved are the joint limits of
executed on a machine running Ubuntu 18.04 with an Intel Core
the robot and the height of the toe, as the system must ensure
i7-12700 K CPU at 4.9 GHz and with 64 GB of RAM. The
the toe clears the step while also ensuring the rear legs can still
motor model described in (34h) is implemented in the Gazebo
reach the terrain for support.
simulation to enable a more realistic execution. The Acceleration
Results from each MPC configuration are shown in Fig. 6 and
environment is simulated in real-time to measure the effect of
quantitatively summarized in Fig. 6(f). The state trajectories are
solve time on stability. The simulations were slowed down by
shown in Fig. 6(b). All of the configurations are able to complete
a factor of 2× for the Step environment and 5× for the Gap
the task and reach the goal at least some of the time, although
environment (with a maximum solve time of 4Δt = 0.12 s) since
the manner in which the task is completed differs. The lack
resolving the constraints in these tasks are still computationally
of constraint information in the Simple configuration and the
intensive.
myopia of the Mixed configuration often result in foot scuffs
when crossing the step, which nearly destabilize the system
and require large control actions to correct. The complex and
A. Acceleration Environment adaptive configurations are able to identify potential constraint
A series of snapshots of the adaptive configuration performing violations caused by the step sooner and react by increasing the
the Acceleration task are shown in Fig. 5(a). Since the peak walking height and rotating the body to more safely navigate the
acceleration of the system occurs at the beginning and end of step [arrows in Fig. 6(b)].
the trajectory, rapidly converging on a feasible solution to the The computational effort of each configuration is shown in
OCP is essential. Crucially, performing this task does not require Fig. 6(c). Unsurprisingly, the Simple configuration is consis-
exact knowledge of the joint constraints and thus computational tently the fastest since its model is the most sparse and it is
efficiency is key. unaware of the nonlinear joint kinematic constraints, whereas the
Results for each configuration are shown in Fig. 5, with state Complex configuration consistently takes the longest, especially
trajectories of candidate trials in Fig. 5(b), solve time data in to find an initial solution and also once it sees the step. Both the
Fig. 5(c), simplicity set data in Fig. 5(d) and 5(e), and quan- mixed and adaptive configurations have intermediate nominal
titative results in Fig. 5(f). The simple configuration exhibits solve times, but differ when the step approaches. The adaptive
the best performance with a 97% increase in top speed over configuration immediately takes much longer to solve the prob-
complex. Mixed performs next best at an 80% increase in top lem as it needs to reason about this new information, but once a
speed, followed by adaptive at a 55% increase in top speed. These valid solution is found it settles back to its nominal solve time
reflect the relative complexity of each configuration—since the as more complex elements are converted back to simple ones.
complex system must reason about extraneous constraints over Meanwhile the mixed formulation only increases in solve time
the entire horizon, it takes longer to solve the problem—shown when the step is within its shorter window of complex elements,
in Fig. 5(c)—and is thus less capable of stabilizing high- and planning the large control forces required to navigate the
acceleration behaviors. The simple configuration conversely step on such short notice causes a large and sustained increase
excels since it is solving a reduced problem. The Mixed and in solve time.
Adaptive systems consist mostly of simple finite elements and The degree of horizon simplification is shown in Fig. 6(d).
thus retain this benefit, although the Adaptive configuration While the fixed-complexity configurations remain uniform for
performance is slightly degraded since more complex elements the entire task, the Adaptive configuration clearly leverages
are added at the beginning of the behavior during periods of additional complexity when encountering the step. However,
high joint torque and velocity, as shown in Fig. 5(e). These even in the worst case around half of the horizon remains
results support Hypothesis 3, which states that the reducing simplified, the effects of which are seen in the lower solve times
model yields improved locomotion performance through more compared with the Complex configuration. These horizons are
efficient computation. visually shown in Fig. 6(e), which shows the prediction horizon
Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4628 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

Fig. 5. Data for Acceleration experiment. The computational efficiency of the simple configuration permits aggressive commands despite model reductions. Both
adaptive and mixed outperform complex because they retain some efficiency. (a) Acceleration environment requires the robot to rapidly move forward 3 m and
come to a rest. Snapshots are equally distributed in time, with increasing opacity corresponding to progression forward in time. (b) Acceleration experiment position
and velocity trajectories show that the additional computation required by the complex configuration significantly reduces its performance. Each curve corresponds
to one trial at the maximum feasible commanded acceleration. (c) Acceleration experiment solve times. Only the adaptive configuration changes with problem
difficulty as joint constraints are activated and deactivated. Chatter at the end corresponds to entering full support phase before coming to rest. (d) Acceleration
experiment simplicity set size, as a percentage of the prediction horizon. The Adaptive configuration shrinks the simplicity set to handle joint constraints but
later expands it when feasible. (e) Acceleration experiment prediction horizons. Horizons at each time are indicated by horizontal slices of finite elements (dots),
where dot color indicates model complexity. Vertical bands correspond to instances where joint constraints (torque or velocity) require additional complexity.
(f) Experimental data for the Acceleration experiment. Bold indicates best observed performance.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4629

Fig. 6. Data for Step environment. The Adaptive configuration is able to leverage admissible reductions for the majority of the behavior while retaining the ability
to react quickly to the kinematic constraints required to navigate the step. Gray shading indicates the step-climbing period in the reference trajectory, with k = 0
corresponding to the beginning of this period. (a) Step environment requires navigating kinematic constraints. Snapshots are shown of the trajectories under the
adaptive configuration. (b) Step environment state trajectories. Complex (red) and adaptive (green) show changes to pitch and yaw (indicated by the arrows) before
the Simple (blue) and Mixed (gold) configurations. (c) Step environment solve times. The increase of four seconds into the behavior corresponds to navigating the
step. (d) Step environment horizon simplicity set size. The adaptive configuration is able to simplify the problem for most of the behavior, and quickly recover
these simplifications once the difficult behavior is resolved. (e) Step environment prediction horizons. Horizons at each time are indicated by horizontal slices of
finite elements (dots), where dot color indicates model complexity. The vertical bands of increased complexity correspond to instances where joints in the front
and back legs approach singularities. (f) Experimental data for the Step environment.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4630 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

at each time with complex and simple elements distinguished TABLE II

SUMMARY OF EXPERIMENTAL RESULTS IN KEY METRICS
by different colors. The Adaptive configuration clearly reacts to
the step—as elements in the terminal region require leaving the
simple manifold, adaptive complexity MPC fills in new elements
with additional complexity. When the terminal region returns to
the simple manifold, the algorithm recognizes this and allows
simple elements back into the horizon. Together, these results
support Hypothesis 2, which states that reductions are frequently
admissible for candidate terrains, and Hypothesis 3, which states
that capturing the complex dynamics and constraints expands the
range of executable tasks.
both model complexity and computational efficiency across each
task. Among the configurations which can solve the difficult Gap
C. Gap Environment environment and avoid stumbling over the Step environment, it
A series of snapshots of the Adaptive configuration navigating performs the best at the highly dynamic Acceleration task due to
the Gap environment are shown in Fig. 7(a). Due to the state- its ability to more quickly react to aggressive commands. This
dependent actuator limits, which reduces peak torque as joint suggests that the ideal scenario for deploying ACMPC rather
velocity increases, the system must accelerate early to ensure than fixed configurations—whether simple or complex—is one
enough velocity to land safely on the other side. Joint kinematics with a high degree of heterogeneity, where nominal operation is
also limit how far forward the legs can reach to prepare for generally expected but occasionally interrupted by the need for
landing, so the system must be aware of these constraints before foresight or reactivity. This describes many common tasks and
takeoff to ensure sufficient controllability. environments for legged locomotion, such as walking over an
Results for each configuration in the Gap environment are open stretch of flat pavement before scrambling over a curb or
shown in Fig. 7 and quantitatively summarized in Fig. 7(f). dodging away from a moving obstacle.
Only the Adaptive configuration has both the efficiency and The computational efficiency afforded by ACMPC could have
model fidelity required to solve this task reliably. The Com- broader benefits for many robotic systems beyond the increase
plex configuration fails in the majority of the trials due to its in agility measured here. Even with modern architectures, many
excessive computational effort, whereas the Simple and Mixed industrial mobile robots—legged systems in particular—have
configurations never succeed because they lack the requisite tight computational, thermal, and power budgets. Algorithms
constraint knowledge and thus do not leap far enough. As in the which use fixed complexity configurations are unable to regulate
Step environment, the configurations which are able to recognize their own draw on these valuable resources, and those that use
constraints near the end of the horizon (Adaptive and Complex) highly complex models such as those discussed in Section II
do so immediately by lowering toward the ground to obtain a may unnecessarily consume these resources during trivial tasks
longer leaping stroke and accelerating forward to ensure enough such as walking over flat terrain. Applying the concept of adap-
velocity to reach the other side of the gap, as shown in Fig. 7(b). tive complexity to those methods would allow resources to be
The solve times for the Gap environment, shown in Fig. 7(c), allocated elsewhere or even invested back into the controller by
demonstrate similar trends as during the Step environment. Both extending the prediction horizon.
Simple and Mixed formulations maintain their performance until This additional performance does not come without limi-
they fail to cross the gap, which quickly causes failed solves. tation. Adaptive complexity MPC is subject to the formula-
Both Complex and Adaptive configurations have periods of tion of the complex system, and in particular the structure of
longer solves to plan the leaping and landing phase, but the its dynamics and constraints. Introducing additional numerical
Adaptive formation is able to recover faster solve times sooner complexity such as hard constraints or nonconvexity can make
due to its ability to convert complex element to simple ones solving the OCP more susceptible to over-constrained prob-
near the end of the leap. This is further supported by the data in lems, local minima, or poor convergence rates, which are then
Fig. 7(f) which shows a threefold reduction in Adaptive solve transferred to the adaptive configuration. This is most notable
times exceeding one timestep compared with the Complex con- in the Step environment, in which the Simple configuration
figuration. The mechanism for this reduction is further illustrated demonstrated remarkable ability to complete the task without
in the simplification percentages shown in Fig. 7(d) and the constraint knowledge due to its numerical robustness. Ongoing
prediction horizons, shown in Fig. 7(e). Even in the worst-case work into well-conditioned OCP formulations such as the QP ap-
portion of the behavior, the Adaptive configuration retains 25% proximation approach of [44] or methods which identify which
simplification of the horizon, and once more feasible elements constraints are most necessary would benefit the approaches
begin entering the horizon the Adaptive configuration can take discussed here by removing the potentially disproportionate
advantage of the reduced complexity to improve solve times. effect of a few complex finite elements.
These results support Hypotheses 2 and 3. Another current drawback of adaptive complexity is robust-
ness to unexpected errors in the simplicity set caused by model
mismatch or disturbances. Introducing additional complexity in
VIII. DISCUSSION the interior of the horizon can degrade the initialization of the
Table II summarizes the key results of the each of the experi- OCP—this is largely why the Mixed configuration outperformed
ments. While no one configuration is optimal for all three tasks, Adaptive in the Acceleration task. This effect could potentially
only the Adaptive configuration is able to retain the benefits of be alleviated by recent approaches which warm-start the OCP

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4631

Fig. 7. Data for Gap environment. The adaptive and complex configurations are able to reason about constraints at the end of the horizon, allowing them to
alter the leap to increase forward velocity and successfully land. Gray shading indicates the leap period in the reference trajectory, with k = 0 corresponding
to the beginning of this period. (a) Gap environment requires navigating both kinematic and dynamic constraints. Snapshots are shown of the trajectories under
the adaptive configuration. (b) Gap environment state trajectories. Complex (red) and adaptive (green) show changes to horizontal velocity and vertical position
(indicated by the arrows) before the Simple (blue) and Mixed (gold) configurations. (c) Gap environment solve times. The increase four seconds into the behavior
corresponds to navigating the gap. The sustained increases for the Simple and Mixed configurations correspond to failed solves after unsuccessful landing. (d)
Gap environment horizon simplicity set size. Similarly to the Step environment, the Adaptive configuration is able to simplify most of the horizon, with the most
complexity occurring when both takeoff and touchdown are within the horizon. (e) Gap environment prediction horizons. Horizons at each time are indicated
by horizontal slices of finite elements (dots), where dot color indicates model complexity. The vertical bands of increased complexity correspond to takeoff and
touchdown, gaps indicate large solve times. (f) Experimental data for the Gap environment.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4632 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

with experiential data [60], or possibly avoided by applying ro- by (23). We claim this sequence is feasible for ACOCP(xck+1 )
bust MPC techniques [61]. Conversely, one approach employed solved at successor state xck+1 := facmpcc
(xck ).
∗a ∗a
in this work to protect against this effect was to avoid re-adding First, the controls u1 , . . . , uN a −1 , which are elements of
elements to the simplicity set once they had been removed. While u∗a (xck ) which was a solution to (11), all lie in Zfa by Propo-
this reduced the instances of interior elements unexpectedly sition 2. It follows from Assumption 2 (since xT ∈ XTc ) that
requiring additional complexity, it resulted in overly conserva- uT (xT ) ∈ Zfc , and thus every element of ũa (xck ) satisfies the
tive simplicity sets—many elements which initially violated the control constraint of (11).
admissibility conditions later satisfied them once the controller Next, we consider the state constraint. By Proposition 1 the
took corrective actions. It may be possible to safely generate state sequence resulting from initial state facmpc c
(xck ) and control
more optimistic simplicity sets, which could enable real-time a c a a a a
sequence ũ (xk ) is x̃ := [x̃0 , x̃1 , . . . , x̃N a ] where
deployment for more difficult tasks.
In addition, many hierarchical systems leverage reduced- x̃ai = x∗a a
i+1 , i = 0, . . . , N − 1 (41)
order models to generate reference trajectories entirely in the
simple system, making infeasible references highly relevant. In- x̃aN a = fucT (xT ) (42)
vestigations into adapting model complexity to handle infeasible and x̃a0 = x∗a c c ∗l c c
1 = f (xk , u0 ) = facmpc (xk ). By Proposition 2,
references such as [62] or other ways to satisfy Assumption 4 ∗a
the states x1 , . . . , xT satisfy the state constraint. Since xT ∈
without requiring a reference would be crucial to expand this
XTc , Assumption 2 implies that fucT (xT ) ∈ XTc ⊂ XN a , so that
work to a broader class of systems. However, it not uncommon
every element of the state sequence x̃a satisfies the state
to define a simple system with the same dynamics as the complex
system but with simpler constraints. This would permit an iden- constraint, and the new terminal state x̃aN a = fucT (xT ) sat-
tity operator for the lifting function without need for a reference, isfies the stability constraint. Hence, ũa (xck ) is feasible for
c
yielding a simpler OCP, which only evaluates constraints where ACOCP(facmpc (xck )) and facmpc
c
(xck ) ∈ XN a .
necessary. This approach is akin to lazy strategies used in motion The proof of Proposition 5 shows that the OCP cost function
planning [63]. decreases along the closed-loop system by leveraging shared
terms in the solutions along with the observation that the cost
of a state in the simple system is equal to the cost of that state
IX. CONCLUSION lifted into the complex system.
This work presents a formulation of ACMPC, which actively Proof (Proposition 5): The sequence pairs (u∗a (xck ),
identifies regions where dynamics and constraints can be simpli- ũ (xk )) and (x∗a , x̃a ) have common elements and thus the cost
a c

fied without compromising the feasibility or stability of the origi- sequence can be simplified as follows:
nal system. Analysis of the proposed approach demonstrates that
under key conditions these simplifications do not compromise VN∗aa (facmpc
c
(xck )) − VN∗aa (xck )
the stability properties of the original system, and can enable new ≤ VNa a (xck+1 , ũa (xc )) − VNa a (xck , u∗a (xck ))
behaviors by acting quickly to perform agile motions or looking
further into the future to execute behaviors. These advantages = (VIc (xT , uT (xT )) + VT (fuT (xT )))
are demonstrated on a simulated quadrupedal robot performing
− VIc (xck , ucacmpc (xck )) + VT (xT )
agile behaviors with challenging environmental constraints, and
in particular expanding the leaping capability through receding + VIc (xck+1 , u∗a a c ∗a
1 ) − VI (xk+1 , u1 ) .
horizon execution with knowledge of joint constraints. (43)
While the MPC formulation presented here was primarily
evaluated in locomotion applications, future work could in- From the definition of the adaptive cost in (12) and Assump-
vestigate its applicability to other domains that employ hier- tion 4, VIc (xck+1 , u∗a a c ∗a c
1 ) − VI (xk+1 , u1 ) ≥ 0. Since xT ∈ XT ,
archical structures, such as manipulation. For example, often in Assumption 2 implies
manipulation settings the internal joints of the manipulator are
neglected and planning is primarily conducted in the space of VT (fuT (xT )) − VT (xT ) ≤ −VIc (xT , uT (xT )). (44)
object motions and forces. ACMPC would enable an efficient Hence, every term on the right-hand side of (43) is positive
handling of manipulator kinematics only when necessary so that semidefinite, and thus (25) is satisfied for every xck ∈ XN a . 
the system can respect these constraints while largely retaining The proof of Theorem 6 follows from the prior propositions:
the benefits of improved efficiency, including faster reactions to
Proof (Theorem 6): The inequalities in (26a) follow from
unexpected object motion or longer planning horizons.
the structure of the value function defined in Assumption 2.
The descent inequality in (26b) is given by the structure of the
APPENDIX
value function along with Proposition 5. Asymptotic stability of
A. Proofs Required for Adaptive Complexity MPC Stability the origin with region of attraction XN a follows from standard
Lyapunov theory [51, Appendix B].
The proof of Proposition 4 shows recursive feasibility by
finding a feasible solution to the ACOCP at the successor state.
Proof (Proposition 4): Since xck ∈ XN a , there exists a solu- B. Proof of Admissibility Conditions for Legged System
tion u∗a (xck ) to ACOCP(xck ). Because u∗a (xck ) is a solution of Proof (Lemma 9): The conditions under which the legged
(11), by Proposition 3 the corresponding predicted terminal state system described in Section VI can be admissibly simplified
xT ∈ XTc . Let the successor control sequence ũa (xck ) be defined rely on the feasibility of the reference trajectory. Lemma 9
Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 NORBY et al.: ADAPTIVE COMPLEXITY MODEL PREDICTIVE CONTROL 4633

states that an index within a lifted trajectory of the legged [6] D. F.-Keil, S. L. Herbert, J. F. Fisac, S. Deglurkar, and C. J. Tom-
system can be admissibly reduced if the lifted components of lin, “Planning, fast and slow: A framework for adaptive real-time safe
trajectory planning,” in Proc. IEEE Int. Conf. Robot. Automat., 2018,
the state-control pair qfoot , q̇foot , and ufoot all lie on the trajectory pp. 387–394.
and satisfy the constraints in (34). We proceed by each condition [7] R. J. Full and D. E. Koditschek, “Templates and anchors: Neuromechanical
for admissibility defined in (15), noting that (15a) is trivially hypotheses of legged locomotion on land,” J. Exp. Biol., vol. 202, no. 23,
pp. 3325–3332, 1999.
satisfied by the assumptions of the lemma. [8] T. Libby, A. M. Johnson, E. C. -Siu, R. J. Full, and D. E. Koditschek,
The condition in (15b) requires that constraints would be “Comparative design, scaling, and control of appendages for inertial re-
satisfied if the system were reduced, i.e., ψ † ◦ ψ(zil ) ∈ Z c . By orientation,” IEEE Trans. Robot., vol. 32, no. 6, pp. 1380–1398, Dec. 2016.
[9] J. H. Reif, “Complexity of the mover’s problem and generalizations,” in
the conditions given on the values of zil in the null space of ψ Proc. 20th Annu. Symp. Found. Comput. Sci., 1979, pp. 421–427.
and the definition of ψ † , it follows that ψ † ◦ ψ(zil ) = zil . Since [10] K. Hauser et al., “Motion planning for legged robots on varied terrain,”
zil ∈ Z c , it follows that ψ † ◦ ψ(zil ) ∈ Z c . Int. J. Robot. Res., vol. 27, no. 11–12, pp. 1325–1349, 2008.
[11] K. Mombaur, “Using optimization to create self-stable human-like run-
The condition in (15c) requires that the complex system is ning,” Robotica, vol. 27, no. 3, pp. 321–330, 2009.
exactly anchored by the simple system at that index in the tra- [12] M. Posa, C. Cantu, and R. Tedrake, “A direct method for trajectory
jectory, i.e., ψx† ◦ f s ◦ ψ(zil ) = f c (zil ). We show this by directly optimization of rigid bodies through contact,” Int. J. Robot. Res., vol. 33,
no. 1, pp. 69–81, 2014.
applying the dynamics and mappings in Section VI, dropping [13] H. Dai, A. Valenzuela, and R. Tedrake, “Whole-body motion planning
the index k for simplicity as follows: with centroidal dynamics and full kinematics,” in Proc. IEEE Int. Conf.
Humanoid Robots, 2014, pp. 295–302.
[14] S. Bartoszyk, P. Kasprzak, and D. Belter, “Terrain-aware motion planning
⎡ ⎤ for a walking robot,” in Proc. IEEE Int. Workshop Robot Motion Control,
q̇lin 2017, pp. 29–34.
⎢ Jω (qang )ω ⎥ [15] P. Fernbach, S. Tonneau, A. D. Prete, and M. Taïx, “A kinodynamic
ψx† ◦ f s ◦ ψ(z l ) = ψx† ◦ ⎢ 
⎣ 1 nj ubody,j − g ⎦
⎥ steering-method for legged multi-contact locomotion,” in Proc. IEEE/RSJ
m Int. Conf. Intell. Robots Syst., 2017, pp. 3701–3707.
W (qlin , q̄foot , ω, ubody ) [16] S. Tonneau, A. D. Prete, J. Pettré, C. Park, D. Manocha, and N. Mansard,
⎡ ⎤ “An efficient acyclic contact planner for multiped robots,” IEEE Trans.
q̇lin Robot., vol. 34, no. 3, pp. 586–601, Jun. 2018.
⎢ Jω (qang )ω ⎥ [17] J. Norby and A. M. Johnson, “Fast global motion planning for dynamic
⎢ ⎥
⎢ ˙q̄foot ⎥ legged robots,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2020,
=⎢ 
⎢ 1 nj ubody,j − g ⎥
⎥ pp. 3829–3836.
⎢ m ⎥ [18] A. W. Winkler, C. D. Bellicoso, M. Hutter, and J. Buchli, “Gait and tra-
⎣W (qlin , q̄foot , ω, ubody )⎦ jectory optimization for legged systems through phase-based end-effector
parameterization,” IEEE Robot. Autom. Lett., vol. 3, no. 3, pp. 1560–1567,
ūfoot Jul. 2018.
[19] L. Sentis and O. Khatib, “A whole-body control framework for humanoids
= f c (z l ). operating in human environments,” in Proc. IEEE Int. Conf. Robot. Au-
tomat., 2006, pp. 2641–2648.
[20] S. Kuindersma et al., “Optimization-based locomotion planning, estima-
tion, and control design for the atlas humanoid robot,” Auton. Robots,
Lastly, the condition in (15d) requires that the dynamics at vol. 40, no. 3, pp. 429–455, 2016.
the prior state lead to the manifold, i.e., ψx† ◦ ψ ◦ f c (zi−1 l
)= [21] M. Neunert et al., “Whole-body nonlinear model predictive control
c l l
f (zi−1 ). Since the trajectory z is valid for the system in (31), through contacts for quadrupeds,” IEEE Robot. Autom. Lett., vol. 3, no. 3,
pp. 1458–1465, Jul. 2018.
xli = f c (zi−1
l
). In addition, since ψ † ◦ ψ(zil ) = zil , it follows that [22] M. Zucker et al., “Optimization and learning for rough terrain legged
ψx ◦ ψ(zi ) = xli . Thus, ψx† ◦ ψ ◦ f c (zi−1
† l l
) = f c (zi−1 l
).  locomotion,” Int. J. Robot. Res., vol. 30, no. 2, pp. 175–191, 2011.
[23] E. Plaku, L. E. Kavraki, and M. Y. Vardi, “Motion planning with dynamics
by a synergistic combination of layers of planning,” IEEE Trans. Robot.,
vol. 26, no. 3, pp. 469–482, Jun. 2010.
ACKNOWLEDGMENT [24] D. McConachie, T. Power, P. Mitrano, and D. Berenson, “Learning when
to trust a dynamics model for planning in reduced state spaces,” IEEE
The authors would like to thank Lorenz T. Biegler and Maxim Robot. Autom. Lett., vol. 5, no. 2, pp. 3540–3547, Apr. 2020.
[25] J. Carpentier, R. Budhiraja, and N. Mansard, “Learning feasibility con-
Likhachev for insightful discussions on the theoretical properties straints for multi-contact locomotion of legged robots,” in Proc. Robot.:
of MPC and other motion planning algorithms. Sci. Syst., 2017, p. 9.
[26] M. Kapadia et al., “Multi-domain real-time planning in dynamic en-
vironments,” in Proc. ACM SIGGRAPH/Eurographics Symp. Comput.
Animation, 2013, pp. 115–124.
REFERENCES [27] M. Brandao, M. Fallon, and I. Havoutis, “Multi-controller multi-objective
locomotion planning for legged robots,” in Proc. IEEE/RSJ Int. Conf.
[1] C. D. Bellicoso et al., “Advances in real-world applications for legged Intell. Robots Syst., 2019, pp. 4714–4721.
robots,” J. Field Robot., vol. 35, no. 8, pp. 1311–1326, 2018. [28] H. Zhang, J. Butzke, and M. Likhachev, “Combining global and local
[2] M. Hutter et al., “ANYmal–Toward legged robots for harsh environments,” planning with guarantees on completeness,” in Proc. IEEE Int. Conf.
Adv. Robot., vol. 31, no. 17, pp. 918–931, 2017. Robot. Automat., 2012, pp. 4500–4506.
[3] H. Kolvenbach et al., “Towards autonomous inspection of concrete de- [29] B. Styler and R. Simmons, “Plan-time multi-model switching for mo-
terioration in sewers with legged robots,” J. Field Robot., vol. 37, no. 8, tion planning,” in Proc. Int. Conf. Automated Plan. Scheduling, 2017,
pp. 1314–1327, 2020. pp. 558–566.
[4] D. Kahneman, Thinking, Fast and Slow. New York, NY, USA: Macmillan, [30] A. Dornbush et al., “A single-planner approach to multi-modal humanoid
2011. mobility,” in Proc. IEEE Int. Conf. Robot. Automat., 2018, pp. 4334–4341.
[5] K. Gochev et al., “Path planning with adaptive dimensionality,” in Proc. [31] F. Allgöwer and A. Zheng, Nonlinear Model Predictive Control, vol. 26.
Int. Symp. Combinatorial Search, 2011, pp. 52–59. Basel, Switzerland: Birkhäuser, 2012.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.
 4634 IEEE TRANSACTIONS ON ROBOTICS, VOL. 40, 2024

[32] J. D. Carlo et al., “Dynamic locomotion in the MIT Cheetah 3 through [59] M. H. Raibert, Legged Robots That Balance. Cambridge, MA, USA: MIT
convex model-predictive control,” in Proc. IEEE/RSJ Int. Conf. Intell. Press, 1986.
Robots Syst., 2018, pp. 1–9. [60] N. Mansard et al., “Using a memory of motion to efficiently warm-start a
[33] A. Laurenzi, E. M. Hoffman, and N. G. Tsagarakis, “Quadrupedal walking nonlinear predictive controller,” in Proc. IEEE Int. Conf. Robot. Automat.,
motion and footstep placement through linear model predictive control,” 2018, pp. 2986–2993.
in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2018, pp. 2267–2273. [61] A. Bemporad and M. Morari, “Robust model predictive control: A survey,”
[34] Y. Shi, P. Wang, M. Li, X. Wang, Z. Jiang, and Z. Li, “Model predictive in Robustness in Identification and Control. London, U.K.: Springer, 1999,
control for motion planning of quadrupedal locomotion,” in Proc. IEEE pp. 207–226.
Int. Conf. Adv. Robot. Mechatronics, 2019, pp. 87–92. [62] I. Batkovic, M. Ali, P. Falcone, and M. Zanon, “Model predictive control
[35] J. Liu, P. Zhao, Z. Gan, M. J. -Roberson, and R. Vasudevan, “Leveraging with infeasible reference trajectories,” 2021, arXiv:2109.04846.
the template and anchor framework for safe, online robotic gait design,” [63] R. Bohlin and L. E. Kavraki, “Path planning using lazy PRM,” in Proc.
in Proc. IEEE Int. Conf. Robot. Automat., 2020, pp. 10 869–10 875. IEEE Int. Conf. Robot. Automat., 2000, pp. 521–528.
[36] V. Kurtz, R. R. d. Silva, P. M. Wensing, and H. Lin, “Formal connections
between template and anchor models via approximate simulation,” in Proc.
IEEE-RAS Int. Conf. Humanoid Robots, 2019, pp. 64–71.
[37] A. Pandala et al., “Robust predictive control for quadrupedal locomotion: Joseph Norby (Member, IEEE) received the B.S. de-
Learning to close the gap between reduced-and full-order models,” IEEE gree in mechanical engineering from the University
Robot. Autom. Lett., vol. 7, no. 3, pp. 6622–6629, Jul. 2022. of Notre Dame, Notre Dame, IN, USA, in 2016,
[38] H. Li, R. Frei, and P. M. Wensing, “Model hierarchy predictive control of and the Ph.D. degree in mechanical engineering from
robotic systems,” IEEE Robot. Autom. Lett., vol. 6, vol. 2, pp. 3373–3380, Carnegie Mellon University, Pittsburgh, PA, USA, in
Apr. 2021. 2022.
[39] A. M. Johnson and D. E. Koditschek, “Legged self-manipulation,” IEEE He is currently the Motion Control and Planning
Access, vol. 1, pp. 310–334, 2013. Lead with Apptronik, Austin, TX, USA. His research
[40] Q. Nguyen et al., “Optimized jumping on the MIT Cheetah 3 robot,” in interests include navigation planning, dynamic loco-
Proc. IEEE Int. Conf. Robot. Automat., 2019, pp. 7448–7454. motion, and contact-rich manipulation for humanoid
[41] H. Kolvenbach et al., “Towards jumping locomotion for quadruped robots robots.
on the moon,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2019,
pp. 5459–5466.
[42] M. Chignoli, S. Morozov, and S. Kim, “Rapid and reliable quadruped
motion planning with omnidirectional jumping,” in Proc. IEEE Int. Conf.
Robot. Automat., 2022, pp. 6621–6627. Ardalan Tajbakhsh (Student Member, IEEE) re-
[43] B. Ponton, M. Khadiv, A. Meduri, and L. Righetti, “Efficient multicontact ceived the B.S. degree (Hons.) in mechanical en-
pattern generation with sequential convex approximations of the centroidal gineering from the University of Illinois, Urbana-
dynamics,” IEEE Trans. Robot., vol. 37, no. 5, pp. 1661–1679, Oct. 2021. Champaign, IL, USA, in 2018, and the M.S. de-
[44] R. Grandia et al., “Perceptive locomotion through nonlinear model predic- gree in mechanical engineering with concentration
tive control,” IEEE Trans. Robot., vol. 39, no. 5, pp. 3402–3421, Oct. 2023. in robotics in 2020 from Carnegie Mellon University,
[45] A. Meduri, P. Shah, J. Viereck, M. Khadiv, I. Havoutis, and L. Righetti, Pittsburgh, PA, USA, where he is currently working
“BiConMP: A nonlinear model predictive control framework for whole toward the Ph.D. degree in mechanical engineering.
body motion planning,” IEEE Trans. Robot., vol. 39, no. 2, pp. 905–922, His research interests include scalable motion plan-
Apr. 2023. ning and control for multiagent systems under uncer-
[46] R. Grandia, F. Farshidian, R. Ranftl, and M. Hutter, “Feedback MPC tain real-world conditions.
for torque-controlled legged robots,” in Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst., 2019, pp. 4730–4737.
[47] J.-P. Sleiman, F. Farshidian, M. V. Minniti, and M. Hutter, “A unified MPC
framework for whole-body dynamic locomotion and manipulation,” IEEE
Robot. Autom. Lett., vol. 6, no. 3, pp. 4688–4695, Jul. 2021.
[48] C. Mastalli et al., “Crocoddyl: An efficient and versatile framework for
multi-contact optimal control,” in Proc. IEEE Int. Conf. Robot. Automat., Yanhao Yang (Graduate Student Member, IEEE)
2020, pp. 2536–2542. received the B.E. degree in mechanical engineer-
[49] C. Mastalli et al., “Agile maneuvers in legged robots: A predictive control ing from the South China University of Technology,
approach,” 2022, doi: 10.1109/TRO.2023.3324580. Guangzhou, China, in 2020, and the M.S. degree in
[50] E. Dantec et al., “Whole body model predictive control with a memory mechanical engineering from Carnegie Mellon Uni-
of motion: Experiments on a torque-controlled talos,” in Proc. IEEE Int. versity, Pittsburgh, PA, USA, in 2022. He is currently
Conf. Robot. Automat., 2021, pp. 8202–8208. working toward the Ph.D. degree in robotics with
[51] J. B. Rawlings, D. Q. Mayne, and M. Diehl, Model Predictive Control: Oregon State University, Corvallis, OR, USA.
Theory, Computation, and Design, vol. 2. Madison, WI, USA: Nob Hill His research interests include geometric motion
Publishing, 2017. planning and control for bio-inspired robots.
[52] D. Limon et al., “Input-to-state stability: A unifying framework for robust
model predictive control,” in Nonlinear Model Predictive Control. Berlin,
Germany: Springer, 2009, pp. 1–26.
[53] P. M. Wensing et al., “Optimization-based control for dynamic legged
robots,” 2022, arXiv:2211.11644.
[54] A. Ferramosca, J. B. Rawlings, D. Limón, and E. F. Camacho, “Economic
MPC for a changing economic criterion,” in Proc. IEEE Conf. Decis. Aaron M. Johnson (Senior Member, IEEE) received
Control, 2010, pp. 6131–6136. the B.S. degree in electrical and computer engineering
[55] A. M. Johnson, S. E. Burden, and D. E. Koditschek, “A hybrid systems from Carnegie Mellon University, Pittsburgh, PA,
model for simple manipulation and self-manipulation systems,” Int. J. USA, in 2008, and the Ph.D. degree in electrical and
Robot. Res., vol. 35, no. 11, pp. 1354–1392, Sep. 2016. systems engineering from the University of Pennsyl-
[56] M. Chignoli and P. M. Wensing, “Variational-based optimal control of vania, Philadelphia, PA, USA, in 2014.
underactuated balancing for dynamic quadrupeds,” IEEE Access, vol. 8, He was a Postdoctoral Fellow with Carnegie Mel-
pp. 49 785–49 797, 2020. lon University and the University of Pennsylvania.
[57] J. Norby et al., “Quad-SDK: Full stack software framework for agile He is currently an Associate Professor of mechanical
quadrupedal locomotion,” in Proc. ICRA Workshop Legged Robots, 2022. engineering with Carnegie Mellon University, with
[58] A. Wächter and L. T. Biegler, “On the implementation of an interior-point appointments in the Robotics Institute and Electrical
filter line-search algorithm for large-scale nonlinear programming,” Math. and Computer Engineering Department. His research interests include legged
Program., vol. 106, no. 1, pp. 25–57, 2006. locomotion, hybrid dynamical systems, robust control, and bio-inspired robotics.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 15:51:32 UTC from IEEE Xplore. Restrictions apply.