Hamel’s Equations and

Geometric Mechanics of
rspa.royalsocietypublishing.org
Constrained and Floating
Multibody and Space Systems
Research
Andreas Müller 1
Cite this article: Müller A. 2023
Hamel’s equations and geometric 1 Johannes Kepler University, Linz, Austria

mechanics of constrained and floating

multibody and space systems. Modern geometric approaches to analytical mechanics
arXiv:2305.05412v1 [math.DS] 9 May 2023

20220732. Proc R Soc A 479, 2023. rest on a bundle structure of the configuration
http://dx.doi.org/10.1098/rspa.2022.0732 space. The connection on this bundle allows for
an intrinsic splitting of the reduced Euler-Lagrange
equations. Hamel’s equations, on the other hand,
Article submitted to journal
provide a universal approach to non-holonomic
mechanics in local coordinates. The link between
Hamel’s formulation and geometric approaches in
Subject Areas:
local coordinates has not been discussed sufficiently.
Geometric mechanics, Lie groups,
The reduced Euler-Lagrange equations as well as
computational mechanics, robotics the curvature of the connection, are derived with
Hamel’s original formalism. Intrinsic splitting into
Keywords: Euler-Lagrange and Euler-Poincaré equations, and
Geometric mechanics, Hamel inertial decoupling is achieved by means of the locked
equations, Hamel coefficients, velocity. Various aspects of this method are discussed.
Euler-Poincaré equations, Lagrange
reduction, mechanical connection,
locked velocity, kinematic 1. Introduction
reconstruction, gauge fields, space
Many dynamical systems and controlled multibody
systems systems possess symmetry invariants, and can be
modeled on a principle bundle. The bundle formulation
Author for correspondence: (for Lagrangian systems) was developed in [1,2] as key
concept in geometric mechanics, where the configuration
Andreas Müller
space is regarded as principle bundle Q = G × Q/G with
e-mail: [email protected]
symmetry group G. Central is the notion of a connection
as it allows encoding specific symmetries of the system
[3–6]. The (natural) mechanical connection, deduced from
the system momentum, was introduced in [7–9]. Given a
G-invariant Lagrangian `, in bundle coordinates (Ω α , ṙI )
of a (left) trivialization, where Ω α is the locked velocity,
the dynamics of unconstrained floating-base systems is
governed by the Lagrange-Poincaré equations [3]
d ∂` ∂`  β I 
α
= β
−EIα ṙ + cβ
αλ Ωλ (1.1)
dt ∂Ω ∂Ω
d ∂` ∂` ∂`  β J β

I
− I = β
−BIJ ṙ + EIα Ωα (1.2)
dt ∂ ṙ ∂r ∂Ω
with connection coefficients Aα α
I and its curvature BIJ .

© The Authors. Published by the Royal Society under the terms of the
Creative Commons Attribution License http://creativecommons.org/licenses/
by/4.0/, which permits unrestricted use, provided the original author and
source are credited.
 α
The coefficients EβI = cα λ
βλ AI , and the curvature coefficients 2

α ∂Aα ∂Aα β λ
BIJ = I
− J
± cα
βλ AI AJ (1.3)

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
∂rJ ∂rI
are determined by the structure constants cα βλ of the Lie algebra g of G. The symmetry group
encodes invariances of the Lagrangian. Most prominent examples are floating-base multibody
systems1 , where the motion g of a base body evolves in a subgroup G of the group of Euclidean
motions SE (3), with body-velocity ξ̂ = g −1 ġ ∈ g, and Q/G is the ’joint space’ (which will be
identified with Vn ) with joint coordinates rI . The kinetic energy is then G-invariant, and the
locked velocity is defined in terms of the mechanical connection as Ω α := uα (see Sec. 5), with

uα = ξ α + Aα I I
I (r )ṙ . (1.4)

The curvature vanishes if and only if the constraints defined by momentum conservation,
and thus the Pfaffian system (1.4), is integrable. Similarly for constrained systems, a kinematic
connection was introduced, and the scleronomic constraints are expressed as uα = 0. Mechanical
systems whose spatial motion is constrained are typical examples, where the constraints restrict
the motion of a base body in G. Now the symmetry group accounts for the invariance of the
constraints, and the curvature vanishes if and only if the constraints are integrable. The dynamics
equations for constrained systems are obviously obtained from (1.2) by setting Ω α = 0. Thus,
in both cases, a kinematic relation of the form (1.4) applies, and the curvature appears in the
equations of motion, and is a central object in motion planning and control.
Geometric mechanics provides an intrinsic and coordinate-free framework for modeling,
analysis, and control of finite-dimensional (discrete) as well as infinite-dimensional (continua)
systems. Local coordinate formulations, as the one above, are used for computations, where it
may be necessary to switch between different local coordinate charts. In this context, the fact
that (1.1,1.2) is a specific form of the Hamel equations [10,11] for finite-dimensional systems in
local coordinates deserves recognition, which is the topic of this paper. In geometric mechanics,
Hamel’s equations are now introduced in an elegant modern form, e.g. [12], on the expense
that the relation to the original Hamel equations is lost, however. Moreover, how the geometric
framework and the Hamel formalism are related, and how connection and curvature of the
configuration space (bundle) are related to the Hamel coefficients, is not discussed in the
literature. Without making explicit reference to the original formulation, (1.1,1.2) are referred to
as Hamel equations [8]. Also in robotics and multibody system dynamics, the Euler-Poincaré
equations (1.1) on SE (3) are often interchangeably referred to as Hamel equations or Euler-
Poincaré equations, e.g. [13].
It is shown in this paper that the coefficients in (1.1,1.2) are naturally derived as the
Hamel coefficients. This provides a link between the original Hamel formalism and the
bundle formulation. The explicit derivation also admits consolidating the different coordinate
expressions of local curvature found in the literature, which is crucial for applying the above
equations (see Rem. 5.3). There are also various aspects that need to be taken into account when
using the geometric formalism. One is the concept of locked velocity [3], and the related concept
of average velocity [14,15] that proved to be powerful tools for control of floating-base systems.
It is discussed that there is no frame which can be associated with this velocity whose motion
is a function of g ∈ G and r ∈ Vn , which has consequences for control of floating-base systems.
Another aspect discussed in this paper is that the mechanical connection on Q induced by
the locked velocity intrinsically splits the reduced Euler-Lagrange equations in horizontal and
vertical. As an important consequence, the equations (1.1,1.2) are inertially decoupled, which
is relevant for computational multibody dynamics. Throughout the paper, all constraints are
assumed to be linear (i.e. catastatic) and scleronomic.

1
Throughout the paper, the term floating-base systems covers a large class of mechanical (control) systems characterized by
a base body free to move in space to which further bodies (links) are geometrically connected, e.g. by joints.
 The paper is organized as follows. In Sec. 2, the classical Hamel equations are summarized 3
for unconstrained and constrained systems. Hamel’s formulation is related to the formulation
on a (locally) trivial bundle in Sec. 3. The obtained relations for the Hamel coefficients in local

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
coordinates are the basis for the derivations in all subsequent sections. In Sec. 4, kinematically
constrained systems with symmetries are treated, where the configuration space Q is a principal
α
bundle, and the curvature coefficients BIJ are obtained as the Hamel coefficients in local
bundle coordinates. This principle bundle approach is adopted in Sec. 5 for unconstrained
α
systems with G-invariant Lagrangian, where the coefficients EβI are obtained immediately as the
Hamel coefficients. A rolling sphere and a floating satellite are used to demonstrate application
of the equations. Finally, unconstrained floating-base systems with conserved momentum are
considered in Sec. 6 with a note on geometric phase and pseudo-holonomic motions. Numerical
simulation results for a satellite and a space robot are reported in Sec. 6 and in the supplement [16].
For background material, an excellent introduction to geometric mechanics can be found in the
text books [3,17] and the overview articles [4,9,18]. Relevant concepts from differential geometry
and on bundles can be found in [3,19,20]. For all derivations, Hamel’s original approach is
the point of departure. This relies on coordinates on Q, and it is necessary to introduce local
coordinates also on the symmetry group (yet the final formulation (1.1,1.2) only needs local
coordinates on Q/G with parameter space Vn ). It applies to finite-dimensional systems for which
always exist local coordinates on Q, and canonical coordinates on G such as multibody systems.
This is in contrast to modern geometric mechanics, where globally valid equations are derived
coordinate free, and local coordinates are introduced when needed. However, the coordinate
formulations allow to relate the geometric approach to Hamel’s formulation. It is assumed
throughout the paper that G is a finite-dimensional Lie group possessing local coordinates. Let
the Lie algebra g be isomorphic to the vector space Rn . Then x̂ ∈ g denotes the Lie algebra
element corresponding to the vector x ∈ Rn ∼ = g. Ricci’s summation convention is used, e.g.
BIa uI = I BIa uI implies summation over index I. The notation is summarized in appendix A.
P

2. The Hamel Equations

(a) Boltzmann-Hamel Equations in Quasi-Velocities
The configuration is described in terms of n generalized coordinates q a , a = 1, . . . , n, with
parameter manifold Q = Vn , which serves as configuration space. In case of a multibody system
with nt translation and nr revolute joints, for instance, Vn = T nr × Rnt , with nr -torus T nr and
n = nr + nt . For such systems, the generalized coordinates may only be locally valid. Quasi-
velocity coordinates ua , a = 1, . . . , n, are introduced that are related to the generalized speeds
q̇ a by
ua = Aab q̇ b , q̇ a = Bba ub (2.1)

where Aab and Bba are smooth functions of q. With vectors of generalized coordinates q ∈ Vn and
quasi-velocities u ∈ Rn , these relations are written in matrix form as

u = A (q) q̇, q̇ = B (q) u. (2.2)

It is assumed that q a are valid local coordinates so that A is regular, and B = A−1 . The relation of
ua to the corresponding quasi-coordinates π a , a = 1, . . . , n is described by the differential forms
dπ a = Aab dq b , and the inverse relation by dq a = Bba dπ b . The quasi-coordinates are generalized
coordinates if and only if the differential forms are exact, otherwise ua are called non-holonomic
velocities [21,22], following [11, p. 473], [23, p. 218].
Denote with L (q a , ua ) a Lagrangian in terms of quasi-velocity coordinates ua , which for
mechanical systems is defined as kinetic energy T (q a , ua ) minus potential energy U (q a ). Then
the variational form of the Hamel equations [10,11,22,24] is
∂L dδπ a − δdπ a
   
d ∂L ∂L a
− − Q a δπ + =0 (2.3)
dt ∂ua ∂π a ∂ua dt
where Qa are generalized forces dual to ua . The explicit form of the Boltzmann-Hamel equations 4
d ∂L ∂L ∂L b c
− b Bab + γac u = Qa (2.4)
dt ∂ua ∂ub

rspa.royalsocietypublishing.org Proc R Soc A 0000000

∂q

..........................................................
is obtained after inserting the transitory relation
dδπ a − δdπ a a b
= γcb u δπ c (2.5)
dt
a
in which γcb are the Hamel coefficients defined as
∂Acr ∂Acs
 
c
γab := − Bar Bbs . (2.6)
∂q s ∂q r
The Hamel coefficients vanish identically if and only if (2.1) are integrable, i.e. if ua are holonomic
velocities. The equations (2.4) along with the kinematic equations (2.1) govern the dynamics in the
non-holonomic tangent bundle (i.e. tangent space defined by non-holonomic constraints). They
are referred to as the Boltzmann-Hamel equations (e.g. in [25,26]) as they where (in a very similar
form) presented by Boltzmann in [27,28] and by Hamel in [10,11,29]. It was Hamel, however, who
generalized them to systems an a Lie group [24,29].
The motion of many mechanical systems can be described on a n-dimensional Lie-group G, so
that quasi-velocities belong the corresponding Lie algebra g. If g (t) ∈ G, such quasi-velocities can
be introduced as left- or right-invariant vector fields, û = g −1 ġ or û = ġg −1 , respectively. Let q a
be canonical coordinates on G, and let u ∈ Rn be the vector representation of û ∈ g, then (2.2)
is a map from Rn to Rn ∼ = g. When using the left-invariant definition of quasi-velocities, the
Hamel coefficients (2.6) are identical to the structure constants ccab of the Lie group. This was
already shown by Hamel [24, p. 428] using the transitory relations dδπ a − δdπ a = γcb a
dπ c δπ c .
c c
Using right-invariant quasi-velocities leads to a change of sign: γab = −cab . Typical example for
such quasi-velocities are the angular velocity (where G = SO (3)) or rigid body twists (where
G = SE (3)). Then left-invariance implies body-fixed representation of angular velocity or twists,
and right-invariance implies spatial representation [30,31]. The explicit derivation of the Hamel-
coefficients for SO (3) and SE (3) using the definition of Hamel coefficients can be found in
[32]. Clearly, π a are holonomic coordinates if and only if G is Abelian. It must be emphasized
that canonical coordinates on G are only locally valid in general. This applies in particular to
SO (3), and thus to SE (3), since it is not simply connected, which leads to the well-known
parameterization singularity of rotations.
In summary, the Hamel equations on a Lie group are the forced Euler-Poincaré equations for
Lagrangian L (q a , ua )
d ∂L ∂L ∂L
± cbac b uc = b Bab + Qa (2.7)
dt ∂ua ∂u ∂q
where the positive sign applies to left-invariant, and the negative sign to right-invariant quasi-
velocities. These are the reduced Euler-Lagrange equations for systems whose Lagrangian is (left
or right) invariant under action of a symmetry group G [9,33].

(b) Hamel Equations for Constrained Systems in Quasi-Velocities

The original Hamel equations for constrained systems where presented in [24]. The velocities
q̇ ∈ Rn are now subjected to m̄ Pfaffian constraints, written as uα = 0, α = 1, . . . , m̄, with

uα := Aαa q
a
a
q̇ . (2.8)

It is assumed that the system of Pfaffian constraints is regular, i.e. the m̄ constraints are
independent. Then, δ̄ := n − m̄ independent quasi-velocity coordinates are introduced as

uI = AIa q a q̇ a , I = m̄ + 1, . . . , n


(2.9)

where δ̄ is the differential DOF of the system (also called instantaneous DOF) [34,35]. The overbar
of δ̄ and m̄ indicate that the constraints are generally non-holonomic. If they are integrable, there
are m = m̄ geometric constraints, and δ = n − m = δ̄ is the finite DOF. The n − m̄ independent 5
coordinates are indexed with capital lattin letters I, J, K. The Pfaffian system (2.8) and the
solution (2.9) are summarized as ua = Aab (q a ) q̇ b , as in (2.1) with index set {a} = {α, I}. The

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
independent coordinates are only locally valid. Moreover, the configuration space of a constrained
system is in general not globally a manifold but possesses singularities. This strictly limits the
global validity all coordinate formulations.
Denote with L(q a , ub ) := L(q a , Bba ub ) the Lagrangian in which q̇ a is replaced by ua , by means
of (2.1). The Hamel equations for the constrained system, in terms of the independent velocities
uI , are then obtained from (2.4) as
d ∂L ∂L ∂L a J
− a BIa + γ u = QI , I = m̄ + 1, . . . , n (2.10)
dt ∂uI ∂q ∂ua IJ

where uα , α = 1, . . . m̄ are set to zero after taking the derivatives. The generalized forces are QI =
BIa Qa . The Hamel coefficients in (2.10) are obtained by restricting (2.6) to indices I, J as

∂Aab ∂Aac
 
a
γIJ = − BIb BJc . (2.11)
∂q c ∂q b

The n − m̄ equations (2.10) complemented with the n kinematic equations

q̇ a = BIa q a uI , I = m̄ + 1, . . . , n


(2.12)

govern the system dynamics in terms of state variables (q a , uI ). Not all of the q a may be
independent if the constraints are not completely non-holonomic. Relation (2.12) is obtained from
(2.2) assuming uI are locally valid coordinates on the tangent space and constraints are regular.
Quasi-velocities uI are integrable if and only if γIJ
K
≡ 0. The constraints (2.8), and thus the
∗ ∗ ∗
co-distribution D ⊂ Tq Q with Dq := span (A (q)) defined by the constraints, are integrable
α
(in Pfaffian sense) if and only if γIJ ≡ 0. The constraint distribution D on Q, defined as Dq :=
ker A (q) ⊂ Tq Q, is thus integrable (in Cartan sense) if and only if the Hamel coefficients vanish
and the constraints are regular (rank A is constant). This may not apply to non-regular constraints.

3. Kinematically Constrained Systems on a Trivial Bundle

Many kinematic control problems can be formulated on a trivial bundle. Trivial because there is
a global splitting into independent and dependent velocities. The independent velocities serve as
control inputs. Moreover, many control system are in Chaplygin form, i.e. the kinematic relations
only depend on the independent coordinates.

(a) Constrained Hamel Equations in Terms of Holonomic Velocities

Consider (mechanical) systems described by coordinates q a and their time derivatives q̇ a , rather
then non-holonomic velocities uI , subjected to scleronomic non-holonomic Pfaffian constraints
(2.8). A set of (locally) independent velocity coordinates can be selected.
A particular choice of independent velocities is to use time derivatives of δ̄ := n − m̄
coordinates. To this end, the coordinates are partitioned as q = (sα , rI ) ∈ Vm̄ × Vδ̄ =: Q, where
ṡα , 1, . . . , m̄ are the dependent, and the remaining ṙI , I = m̄ + 1, . . . , n are independent velocity
coordinates, i.e. uI := ṙI . This presumes that ṙI are valid local coordinates on Q. Notice that for
non-holonomic constraints, this dependency does not hold true for the coordinates sα , rI , and Q
serves as n-dimensional configuration space. The constraints (2.8) are then written as

uα := Aα β q
a
β
ṡ + AαI q
a
I
ṙ , α = 1, . . . , m̄ (3.1)
! !
A1 A2 ṡ
, with A1 = Aα α



u = A (q) q̇ = β , A2 = AI . (3.2)
0 I ṙ
where (3.2) resembles the matrix form (2.2). The inverse relation of (3.1) and (3.2) are, respectively, 6
α
ṡ = Bβα q a
β
u + BIα q a
I
ṙ (3.3)

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
! !
ṡ B1 B2
= B (q) u, with B = , with B1 = A−1 −1
1 , B2 = −A1 A2 . (3.4)
ṙ 0 I

With A in (3.2) and B in (3.4), the expression (2.11) gives rise to the Hamel coefficients

∂Aα ∂Aα
 
µ
α
γβλ = − ν
Bβµ Bλν
∂sν ∂sµ
∂Aα ∂Aα
 
α
γIJ = r
− s
BIr BJs (3.5)
∂q s ∂q r
∂Aα ∂Aα ∂Aα
! ! !
∂Aα ∂Aα β ∂Aαµ β µ ∂Aα β β β ∂Aα
= I
− J
+ − BI BJ + I
− BJ + − J
BIβ
∂rJ ∂rI ∂sµ ∂sβ ∂sβ ∂rI ∂rJ ∂sβ

∂Aα ∂Aα ∂Aα ∂Aα ∂Aα ∂Aα

     
µ ν µ µ
α
γβJ = r
− s r s
Bβ BJ = ν
− Bβ BJ + − J
Bβµ , α = 1, . . . , m̄.
∂q s ∂q r ∂sµ ∂sν ∂rJ ∂sβ
K
Since the velocities are integrable, it holds true that γab ≡ 0. The Hamel coefficients vanish if and
only if the constraints are holonomic. The expressions (3.5) will be central throughout the paper
as the individual terms in (3.5) allow deriving the coordinate form of the reduced Euler-Lagrange
equations for systems with symmetry directly from the Hamel formulation.
The Lagrangian is written as L(sα , rI , ṡα , ṙI ) to indicate the coordinate partitioning. As in
(2.10), denote with L(sα , rI , ua ) := L(sα , rI , Bβα uβ + BIα uI , uI ) the Lagrangian with ṙI = uI and
∂L β
ṡα replaced by (3.3). The Hamel equations follow from (2.10). Noting that ∂u∂L
α = B , and
∂ ṡβ α
I I
(with slight abuse of notation) identifying u = ṙ , yields the Hamel equations in independent
holonomic velocities ṙI
d ∂L ∂L ∂L α J
− I + α γIJ ṙ = QI , I = m̄ + 1, . . . , n (3.6)
dt ∂ ṙI ∂r ∂ ṡ
where uα are set to zero, i.e. the solution

ṡα = BIα q a ṙI

(3.7)

of the constraints is imposed, after taking the derivatives. The dynamic equations (3.6) along with
the kinematic equations (3.7) govern the dynamics of the non-holonomically constrained system
in terms of the state (q a , q̇ a ) evolving on the non-holonomic tangent bundle.

(b) Kinematic Constraints in Terms of a Bundle Connection

The kinematic constraints uα = 0 are now formulated with

uα := ṡα + Aα
I q
a
I
ṙ , α = 1, . . . , m̄ (3.8)
−1
where (Aα α I I
I ) := A1 A2 = −B2 , with B2 = (BI ) in (3.3). Combined with u := ṙ , this is written
in matrix form as
! !
I −B2 ṡ
u = Ā (q) q̇ = . (3.9)
0 I ṙ

which possesses the obvious inverse relation ṡα = uα − Aα a I

I (q ) u , analogously to (3.7), and thus
!
I B2
q̇ = B̄ (q) u = u. (3.10)
0 I
Noting the specific structure of (3.9) and (3.10), and that only Aα
I depends on q, the corresponding 7
α α ∂Aα
Hamel coefficients are found from (2.11) as γβJ = −γJβ =− J
∂sβ
, and

rspa.royalsocietypublishing.org Proc R Soc A 0000000

∂Aα ∂Aα ∂Aα ∂Aα

..........................................................
J β
α
γIJ = I
− J
+ A − I
Aβ , α = 1, . . . , m̄; I, J = m̄ + 1, . . . , n. (3.11)
∂rJ ∂rI ∂sβ I ∂sβ J
The configuration space Q = Vm̄ × Vδ̄ is regarded as a trivial bundle2 over the base manifold
V with fiber Vm̄ , and bundle coordinates (sα , rI ) ∈ Vm̄ × Vδ̄ . The horizontal space of this trivial
δ̄

bundle is the constraint distribution, i.e. the vector space of velocities satisfying the constraints.
The homogenous kinematic constraints (3.8) define a connection on this bundle. Writing the
constraints in terms of the Pfaffian forms ω α := uα dt = dsα + Aα I
I dr , a connection is introduced
α ∂
as A = ω ∂sα . This is referred to as an Ehresmann connection [3,18,33] with reference to the
original publication [36], and Aα I are the local coordinates of the connection. Since it arises
from the kinematic constraints, it is called the kinematic connection [18]. The connection relates
(independent) motions in the base manifold Vδ̄ to motions in the fiber. Whether this connection
α
(i.e. the constraints) is holonomic is revealed by its curvature, denoted BIJ . Moreover, the
curvature of the kinematic connection plays a key role in the control of constrained mechanical
systems [3,18,37] as well as in locomotion planning [38] as it encodes how motions in the base
manifold generate motions in the fiber. On the trivial vector bundle, the Lie bracket in the
∂Aα β ∂Aα β
curvature (1.3) is the Lie bracket [AJ , AI ]α = ∂s β AI − ∂sβ AJ of
J I
vector fields AI , AJ on Q,
α ∂Aα ∂AJα
so that the local curvature is BIJ
= − I
∂r J ∂r I
+ [AJ , AI ]α [18, p.
32], [3, p. 108]. The Hamel
coefficients (3.11) are thus clearly related to the coordinate form of the curvature as follows.

Proposition 3.1. The Hamel coefficients (3.11) are identical to the curvature components of the kinematic
connection A, in bundle coordinates (sI , rα ), induced by the constraints with (3.8), i.e. BIJ
α α
= γIJ .

Although local coordinates are used in this paper, it should be mentioned that the curvature
of a connection A is its covariant derivative, written coordinate-free as B (X, Y ) = dA (X, Y ) −
[A (X) , A (Y )], with horizontal vector fields X, Y , i.e. Bα (X, Y ) = BIJ
α
X I Y J = γIJ
α
XI Y J .

Remark 3.1. The constraints are holonomic if and only if the kinematic connection is flat, i.e. the
α
components BIJ of the curvature 2-form vanish identically. In this case, Vδ̄ serves as configuration space.
Constraints are said to be in Chaplygin form if A = A (r), referring to Chaplygin’s publications [39,40]. In
α ∂Aα ∂Aα
this case, the Hamel-coefficients (2.10) reduce to γIJ = ∂rJI − ∂rIJ , which implies the obvious condition
∂Aα ∂Aα
I
∂r J
≡ J
∂r I
for integrability of dsα + Aα J I
I (r )dr = 0.

(c) Hamel Equations on a Trivial Bundle, Lagrange–d’Alembert equations

The Hamel coefficients can now be identified with the components of the local curvature. Then
the Hamel equations of the constrained system in holonomic bundle coordinates (rI , sα ) follow
from (3.6) as
d ∂L ∂L ∂L α J
− I + α BIJ ṙ = QI , I = m̄ + 1, . . . , n (3.12)
dt ∂ ṙI ∂r ∂ ṡ
with QI = AaI Qa , in which uα is set to zero, and uI is replaced by ṙI . The dynamic equations
are written in terms of the coordinates rI on the base manifold of the bundle. The remaining
coordinates, the fiber coordinates, are obtained as solution of

ṡα = −AαI q
a
I
ṙ . (3.13)

The constrained dynamics is governed by the Hamel equations (3.12) along with the kinematic
equations (3.13). The equations (3.12) are obtained as constrained Lagrange–d’Alembert
2
Q = M × F is a trivial bundle if it can be written as Cartesian product of a manifold M and F , and if there is a projection
π : Q → M [19]. F is called the standard fiber. For the considered systems, the base manifold is the coordinate subspace Vδ̄
corresponding to the independent velocities, and the fiber is the subspace Vm̄ corresponding to dependent velocities.
equations with variations satisfying the constraints 0 = δsα + Aα I I
I (r )δr [3,8,18], which is a 8
particular form of Hamel’s equations when using a connection to introduce constraints.

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
Remark 3.2. Introduce the constrained Lagrangian Lc (sα , rI , ṙI ) := L(sα , rI , −Aα I I
I ṙ , ṙ ) =
α I α I I
L(s , r , u := 0, u := ṙ ), i.e. the Lagrangian with the constraints resolved. The Hamel equations
(3.12) attain the instructive form
d ∂Lc ∂Lc ∂Lc ∂L α J
− + α Aα
I + γ ṙ = QI , I = m̄ + 1, . . . , n, (3.14)
dt ∂ ṙI ∂rI ∂s ∂ ṡα IJ
which reveals the consequence of non-holonomicity of the constraints. They were reported in [37, p. 326]
α α
(setting γIJ = BIJ ), and in a similar form for constraints independent of sI by Chaplygin [22,39]. The
Hamel coefficients reveal the consequence of non-holonomic constraints. Clearly, if the constraints are
completely holonomic, these are the classical Lagrange equations. A direct calculation shows that (3.14)
can be written in the form
 
d ∂L ∂L a d ∂L ∂L
− + A I − = AaI Qa , I = m̄ + 1, . . . , n (3.15)
dt ∂ ṙI ∂rI dt ∂ ṡα ∂sα
which have been reported by Voronets [41,42], and are referred to as Voronets equations [25,43].

4. Kinematically Constrained Mechanical Systems with Symmetry

Many kinematically constrained systems possess principal symmetries in the sense that the
kinematic constraints are invariant under the action of a symmetry group G. The configuration
space can then be regarded as a principal bundle3 . Moreover, this is a trivial principal bundle
when the kinematic constraints do not depend on group variables. Examples are mobile platforms
and manipulators, or locomotion systems, where G is often a subgroup of SE (3), the group of
rigid body (i.e. Euclidean) motions, and g ∈ G describes the motion of a base body. The governing
equations are the reduced Euler-Lagrange equations on the principle bundle.

(a) Constraints in Terms of a Connection on a Trivial Principal Bundle

The (non-holonomic) constraints are assumed to be invariant under the action of a Lie group G.
Chaplygin systems are included as special case with Abelian symmetry group. The configuration
space of the system is regarded as a trivial principal bundle Q = G × Vδ̄ over the base manifold
Vδ̄ with fiber G. The dimension of G is assumed to be equal to the number of constraints,
and is denoted with m̄ (to be consistent with the preceding section). Typically, fiber elements
g ∈ G represent the overall configuration of the system in ambient space, and are often called
’body coordinates’ (or ’rigid coordinates’). The coordinates r = (rI ) ∈ Vδ̄ represent the internal
shape, and are called ’shape coordinates’ (or ’internal variables’), and Vδ̄ is called the shape space.
For multibody systems, r is the vector of joint variables. Notice that rI are only locally valid
coordinates in general. The case when the number of constraints is less then the dimension of the
symmetry group G has been addressed for motion planning of non-holonomically constrained
mechanical control systems in [44].
Kinematic constraints that are left-invariant under actions of G are expressed as Pfaffian
system uα = 0, α = 1, . . . , m̄, with

uα := ξ α + Aα I I
I (r )ṙ (4.1)

where ξ̂ = g −1 ġ ∈ g, and ξ = (ξ α ) ∈ Rn are the fiber coordinates in a left-trivialization. More

precisely, for rigid body systems (where G = SE (3)), ξ is the velocity (also called twist) of
3
A space Q = G × Q/G is a principle bundle, where G is a Lie group, with Lie algebra g, acting free and proper on Q,
equipped with a projection π : Q → Q/G [20]. In a local trivialization, the base space Q/G can be identified with a manifold
B so that Q = G × B. If this splitting is globally valid, Q is a trivial principle bundle. In a local trivialization, with local
coordinates (ξ α , ṙ I ), B will be identified with Vδ̄ . For a given r ∈ Vδ̄ , π −1 (r) is the fiber over r. For mechanical systems, r
describes the internal configuration (shape) of the system, and fiber elements g ∈ G represent the pose of a base body.
a reference body in ’body-fixed’ representation. The constraints (4.1) give rise to a principal 9
connection on the trivial principal bundle, denoting Adr = (Aα I
I dr ),

rspa.royalsocietypublishing.org Proc R Soc A 0000000

Akin = Adg (g −1 dg + Adr) = dgg −1 + Adg (Adr)

..........................................................
(4.2)

so that the horizontal subspace of the connection is the space of velocities satisfying the
constraints, and Akin is a g-valued one-form (which may be considered as a special type of
Ehresmann connection) called the kinematic connection as it arises from (4.1) by requiring it to be
G-equivariant [3]. The name stems from the fact that it relates base and fiber motions according to
the kinematic constraints. It is sufficient to use the local connection form Aα
I in (4.1) as it encodes
all relevant information. The symbol Akin is used to distinguish it from the coefficients of the
local form Aα I . Next, the Hamel coefficients are derived and are identified as the coefficients of
the curvature of the kinematic connection.

(b) The Hamel Coefficients and the Kinematic Connection

The dynamics of the constrained system is governed by the reduced Lagrange-d’Alembert-
Poincaré equations, which have been derived from the variational principle [9,45]. To derive them
as the constrained Hamel equations necessitates the corresponding Hamel coefficients.

Lemma 4.1. The non-vanishing Hamel coefficients in (2.10) for the system subjected to left G-invariant
constraints are
α ∂Aα ∂Aα λ µ
γIJ = I
J
− J
+ cα
λµ AI AJ , α = 1, . . . , m̄. (4.3)
∂r ∂rI

Proof. In order to apply the original definition (2.6) of the Hamel coefficients, local canonical
coordinates sα , α = 1, . . . , m̄ are introduced on G. The fiber coordinates are then expressed as
ξ α = Aα α β α α α β
β (s ) ṡ , with inverse relation ṡ = Bβ (s ) ξ , and (4.1) is written as

uα = Aα
β s
α
β
ṡ + Aα J I
I (r )ṙ . (4.4)

This resembles the relation (3.1) with Aα α α α α J

β = Aβ (s ) and AI = AI (r ). The inverse relation is

ṡα = Bβα sa uβ − Bβα sa Aβ J I

I (r )u (4.5)

with Bβα defined in (3.4). Noting that AIa = const, the only non-zero Hamel coefficients for the
α
constrained system are γIJ . They are immediately found from (3.5), by replacing Aα J
I (r ) with
α J α α a β J
AI (r ), and BI with −Bβ (s ) AI (r ), as

∂Aα ∂Aα ∂Aα ∂Aα

 
α
γIJ = I
− J
+ δ
− λ
Bγδ Bµλ AγI Aµ
J. (4.6)
∂rJ ∂rI ∂sλ ∂sδ
 ∂Aα
∂Aα

It was already shown by Hamel [24, p. 428] that the terms γβλ α
= ∂sνµ − ∂sµν Bβµ Bλν = cα βλ are
the structure constants of G. A more recent reference is [46, p. 301]. This was derived explicitly
in [32] for SO (3) andSE (3) using the original definition of Hamel coefficients.

The expressions (4.3) are the coefficients of the local curvature of the connection Akin on the
principal bundle possibly up to a change of sign [19,47] (for the sign convention see Rem. 5.3). In
context of geometric mechanics, this can be stated as follows [18, notice the correction on p. 44].

Proposition 4.1. The Hamel coefficients (4.3) for left-invariant kinematic constraints (4.1) are the
components of the curvature (1.3) of the kinematic connection defined in bundle coordinates (ξ α , ṙI ) on
α α
the corresponding left-trivialized principal bundle: BIJ = γIJ . Noting that the curvature coefficients are
identical to the Hamel coefficients, the connection is flat, if and only if the constraints are holonomic.
 Also the curvature on the principle bundle can be defined coordinate-free as covariant 10
derivative B (X, Y ) = dA (X, Y ) − [A (X) , A (Y )], with horizontal vector fields X, Y , now with
Lie bracket on g. In local coordinates, it is Bα (X, Y ) = γIJ
α
XI Y J .

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
−1
Remark 4.1. If right trivialization  ξ̂ = ġg ∈ g are right invariant vector fields (e.g. using
is used, i.e.
∂Aα ∂Aα

α
spatial velocities), then γβλ = µ
∂sν − ν
∂sµ Bβµ Bλν = −cα
βλ , and the Hamel coefficients (4.3) are

α ∂Aα ∂Aα λ µ
γIJ = I
J
− J
− cα
λµ AI AJ , α = 1, . . . , m̄. (4.7)
∂r ∂rI

Remark 4.2. The constraints (4.1) give rise to a kinematic control system on G that can be written as
ξ α = −Aα I I
I (r )ṙ . Such control systems with symmetry were addressed in various publications, e.g. [38,
44,48,49]. Controllability of this driftless control problem is encoded in the control vector fields Aα I and the
distribution on the fiber defined by them. Written as ġ = −gAdṙ shows that the connection describes how a
path in shape space is lifted to a path in the group (called the horizontal lift of the curve), which is the basis
for kinematic control. The net change in the group variable as result of the horizontal lift of a closed curve
in shape space is the holonomy (in this context called the geometric phase). The latter can be related to the
curvature of the kinematic connection. Let φ be a closed path in Vδ̄ . The geometric phase is then found as
I Z
g (φ) = − gAdr = − B + hot. (4.8)
φ Φ

using ġ = g ξ̂ = −gAṙ, where Φ is the area enclosed by the path in Vδ̄ , and B = (BIJ α
drI drJ ). If the
constraints are holonomic, i.e. the curvature vanishes, the geometric phase is zero. This is an important
relation for locomotion planning, where the closed path φ represents a gait, and the aim is to maximize the
net motion in G generated by a gait [38]. The geometric phase further discussed in Sec. 6.

(c) The Hamel Equations, Lagrange–d’Alembert–Poincaré equations

The Lagrangian L(g, rI , ġ, ṙI ), defined on the configuration space Q, is assumed to be (left
or right) G-invariant so that the reduced Lagrangian l(rI , ξ α , ṙI ) can be introduced on the
corresponding bundle trivialization. Assume that the motion in G is completely determined by
the motion in the base manifold, i.e. the group orbits complement the constraints. This is called
the ’principal kinematic case’ [3,18] since then there is no momentum equation left on the fibre.
The Pfaffian equations (4.1) can be resolved as ξ α = uα − Aα I
I ṙ . In the Hamel formalism, the
α α
u are regarded as intermediate coordinates, and setting u = 0 yields the constraint solution.
Denote with `(rI , uα , uI ) := l(rI , ξ α := uα − Aα I I I α
I ṙ , ṙ := u ) the Lagrangian in terms of u in
I I
(4.1) and ṙ := u . The Hamel equations (2.10) for the constrained system are expressed using
proposition 4.1 as
d ∂` ∂` ∂l β J
− I = B ṙ + QI , I = m̄ + 1, . . . , n (4.9)
dt ∂ ṙI ∂r ∂uβ IJ
where uα is set to zero, and uI = ṙI . These are indeed equations (1.2) when setting Ω α = uα = 0.
The velocities in the fiber are obtained with the kinematic connection as ξ α = −Aα I
I (r) ṙ , which is
the principle bundle equivalent of (3.13). In geometric mechanics, equations (4.9) are obtained as
Lagrange–d’Alembert–Poincaré equations with variations satisfying the constraints uα = 0 with
(4.1). The motion g (t) in the fiber G is obtained by solving the kinematic reconstruction equations

ġ = g ξ̂ (left trivialization) or ġ = ξ̂g (right trivialization) (4.10)

where ξ = −Aṙ. The Hamel equations (4.9) along with the reconstruction equations (4.10) govern
the dynamics of the constrained system on the principle bundle Q. Chaplygin systems (Rem. 3.1)
∂Aα ∂Aα
are special cases with Abelian symmetry group G = Rm̄ , where BIJ
α
= J
∂r I
− I
∂r J
.
Remark 4.3. Equations (4.10) are the Poisson equations on G. They are known in context of rigid body 11
kinematics as the left- and right Poisson-Darboux equations, referring to [50], or as the generalized Poisson-
Darboux equations [51]. In order to solve these differential equations on G, g is expressed as the exponential

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
of a η (t) ∈ g, and (4.10) are replaced by the following differential equations on g

η̂˙ = dexp−1 −1
−η̂ (ξ̂) = −dexp−η̂ (Aṙ), with g = g0 exp(η̂) (left trivialization) (4.11)

η̂˙ = dexp−1 −1
η̂ (ξ̂) = −dexpη̂ (Aṙ), with g = exp(η̂)g0 (right trivialization)

with initial value g0 ∈ G, where η̂ (t) ∈ g represents a local canonical parameterization of G, and dexp is
the right-trivialized differential of the exp map on G. The latter is defined by ġg −1 = dexpη̂ (η̂) assuming
g = exp(η̂)g0 . This replacement is a key step in Lie group integration schemes [52–54]. For many Lie
groups relevant to solid mechanics, there are closed from expressions for the dexp map, in particular for
SO (3) and SE (3) [55]. The maps dexp and dexp−1 can be evaluated using truncated series expansions.

(d) Example: Homogenous ball rolling without slipping or spinning

As a simple example, consider a ball on a horizontal plane that is subjected to pure rolling, i.e.
in addition to the rolling-without-slipping constraint it is further constrained so that it cannot
spin about its instantaneous vertical axis, which is parallel to the plane normal. Its configuration
is described by its orientation and the location of its point of contact with the plane. The
configuration space is the bundle SO (3) × R2 over M = R2 . The motion in the fiber G = SO (3)
(the orientation) is completely determined by the motion of the contact point so that this example
is a ’principal kinematic case’. Since the constraints and the potential energy (gravity) are right
G-invariant, the kinetic energy is bi-invariant, the configuration space is regarded as a right-
trivialized trivial principal bundle. It is assumed that the center of mass (COM) of the ball is
at its geometric center. Denote with ρ ∈ R3 the vector from contact point to the COM of the ball,
with p ∈ R3 the position of contact point, and R ∈ SO (3) describes the rotation of the body-fixed
frame Fb relative to the inertial frame F0 . The angular velocity ω ∈ R3 ∼ = so (3) of the ball in
spatial representation is defined by ω̂ = ṘRT ∈ so (3). All vectors are expressed in inertial frame
F0 . This frame is introduced such that its 3-axis is parallel and directed along the plane normal,
so that ρ = (0, 0, R)T , where R is the ball radius.
The rolling condition is 0 = ṗ + ρω,
e which can be transformed to 0 = ρ eṗ + R2 ω − (ρρT )ω.
Here ρ e is the skew symmetric matrix so that the cross product of vector x and y is x ey = x × y
that is written in components as εijk xj y k , with the Levi-Civita symbol εijk . The non-spinning
condition is 0 = ρT ω, which amounts to the constraint ω 3 = 0, and the rolling condition simplifies
to 0 = ω + R12 ρ eṗ. Since ρ is along the plane normal, the rolling condition is 0 = R12 εαJI ρJ =
1 4 5 1 2 2
R εα3I . The ’shape coordinates’ r = (r , r ) = (p , p ) ∈ R are the coordinates of the contact

3 2

b
1

3
2 p

Figure 1. Ball rolling without spinning about the vertical axis.

location in the plane. For sake of compactness, denote I¯ = I − 3. Local bundle coordinates are 12
¯
ω α , α = 1, 2, 3 and ṙI = ṗI , I = 4, 5. The m̄ = 3 kinematic constraints are 0 = uα , α = 1, 2, 3,

rspa.royalsocietypublishing.org Proc R Soc A 0000000

(
1

..........................................................
α α α I
u := ω + AI ṙ , AI := α R εα3I¯, α = 1, 2 (4.12)
0, α=3 .
α β µ
The Hamel coefficients are γIJ = −cα
βµ AI AI , I, J = 4, 5 (negative sign is due to right-
β µ
trivialization), which is only non-zero for α = 3 as β, µ = 1, 2. Evaluation yields cα
βµ AI AJ =
1 α β I ¯
1
ε ε ε
R4 αβµ βr I¯ µsJ¯
ρr ρs = 1
(δ δ − δµI¯δαr )εµsJ¯ρr ρs
R4 µr αI¯
= − R14 ρα εIs s
¯ J¯ρ = − R2 ρ AJ δβ , thus
(
α 0, α = 1, 2
γIJ = 1 I¯ (4.13)
R AJ , α=3 .

In summary, the non-zero coefficients are A24 = −A15 = R 1

and B54 3 3
= −B45 3
= γ54 = R12 .
The Lagrangian is identical to the kinetic energy (potential energy does not affect the motion
of the ball). The right-reduced Lagrangian is l(ω α , ṙI ) = m I J 1 α β
2 δIJ ṙ ṙ + 2 Θαβ ω ω , where Θαβ is
the inertia tensor of the homogenous ball w.r.t. its COM expressed in the spatial inertial frame F0 ,
and m is its mass. The Lagrangian in terms of uI = ṙI and uα in (4.12) is
m 1 β J
`(uα , uI ) = l(ω α := uα − Aα I I I
J u , ṙ := u ) =δ uI uJ + Θαβ (uα − Aα I β
I u )(u − AJ u ).
2 IJ 2
(4.14)
¯
The Hamel equations (4.9), with QI = 0, are found as (with rI = pI )
β J 1 ¯
(mδIJ − Θαβ Aα
I AJ )r̈ − Θ Aβ ṙK AIJ ṙJ = 0, I = 4, 5 (4.15)
R 3β K
where α, β ∈ {1, 2}. The second term in (4.15) is the force due to the gyroscopic torque caused by
the pure rolling. To see this, notice that π3 = Θ3β Aβ K
K ṙ is the vertical component of the angular
momentum πα = Θαβ ω due to the rolling (note ω = 0), and the gyroscopic torque is εαβλ ω β πλ .
β 3

The force is obtained as cross product of this torque with ρ, noting that π1 = π3 = 0 and ω3 = 0.

Remark 4.4. The principal kinematic case can be extended by allowing the Lagrangian to be invariant
under actions of a subgroup of the symmetry group G of the constraints, i.e. infinitesimal generators of
this subgroup lie in the constraint distribution. Then the system is said to possess horizontal symmetries
(relative to the constraints) [3,4,18]. In this case, the kinematic connection along with a connection
accounting for the horizontal symmetry can be introduced. As an example for mechanical systems with
horizontal symmetries, a ball moving on a plane, i.e. a rolling ball that is free to spin about the plane
normal, was used in [4,18]. The horizontal symmetry is then due to the momentum being invariant under
the subgroup SO (2) ⊂ SO (3) of rotations about the plane normal. The m̄ rolling constraints are the 1-
and 2-component of the condition 0 = ṙ + ρωe above. The invariant momentum is the 3-component of Θω.

5. Floating-Base Mechanical Systems with Symmetries

If the Lagrangian L : T Q → R is G-invariant and can be written as L(g, rI , ġ, ṙI ), the n-
dimensional configuration space is regarded as a principal bundle, which admits the local
trivialization Q = G × Q/G. This parallels the formulation of constrained systems in Sec. 4,
where G was the symmetry group of the constraints. The quotient space Q/G is again the shape
space, which can locally be identified with Vδ̄ . This setting applies to many mechanical systems
possessing symmetry invariants. Typical examples are floating multibody systems (humanoids,
areal vehicles, space robots), where the fiber G is a subgroup of SE (3) representing the
overall spatial motion, and Q/G ∼ = Vδ̄ represents the internal shape. An early problem that was
considered as a controlled floating-base system within this framework is the falling cat [56–61].
As an example, consider a human body (or humanoid robot) model. The spatial hip motion
is described by g ∈ G = SE (3), and the motion of body segments (body shape) relative to the
hip is described by the joint variables (angles) r ∈ Vδ̄ . The kinetic energy is invariant under left
G-actions, while the potential energy is invariant under rotations about the vector of gravity. 13
Thus, with the obvious choice G = SE (3), and with the Lagrangian defined as kinetic minus
potential energy, Q is not a principal bundle. This situation is referred to as symmetry breaking.

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
However, for most mechanical systems, a G-invariant Lagrangian can be defined, e.g. restricting it
to the kinetic energy and including potential forces separately. Moreover, for discrete mechanical
systems, this bundle is often trivial, so that the configuration space is Q = G × M , with shape
space M , which will be identified with Vδ̄ in the following, as in Sec. 4. This is naturally so for
floating base multibody systems. It is assumed in the following, that Q is a principal bundle
, and a local trivialization is assumed. In a local bundle trivialization, the Lagrangian is then
written as l(rI , ξ α , ṙI ) = T (rI , ξ α , ṙI ) − V (rI ) (Note that it is still possible to introduce a potential
depending on shape variables rI , e.g. accounting for compliant elements).

(a) Boltzmann-Hamel Equations Reduced Euler–Lagrange Equations

The Boltzmann-Hamel equations are expressed in terms of the velocity coordinates (ua ) =
(ξ α , ṙI ), where ξ α = Aα α β β
β (s ) ṡ are expressed in terms of canonical coordinates s on G (as in Sec.
I
4(b)). The Hamel coefficients γbc ≡ 0 follow immediately noting that ṙI are integrable, and γbc
α
≡0
α I α α
follows from (2.6) noting that ξ are independent of r . The remaining coefficients γβδ = ±cβδ are
the structure constants of G (lemma 4.1). The Boltzmann-Hamel equations (2.4) to the Lagrangian
l(rI , ξ α , ṙI ) are thus the well-known equations [9,33]

d ∂l ∂l
± β cβ ξλ = 0 (5.1)
dt ∂ξ α ∂ξ αλ
d ∂l ∂l
− I =0 (5.2)
dt ∂ ṙI ∂r

where in (5.1) the positive sign holds for a left-trivialization (ξˆ = g −1 ġ body velocities), and the
negative sign when right-trivialization is used (ξˆ = ġg −1 spatial velocities). The base motion is
reconstructed with the respective kinematic equation in (4.10).
The above equations obviously split into the Euler-Poincaré equations (5.1) and Euler-
Lagrange equations (5.2), where the first equation (5.1) can also be written as Lie-Poisson
∂l
equations on G when expressed with momentum Πα = ∂ξ α . While here they have been solely

derived with the Hamel formalism, they were derived as reduced Euler-Lagrange equations [9,33]
from a variational principle on Q. The involved variations are not intrinsic in the sense that they
are not split into variations in the fiber and the base manifold, respectively. Such a splitting leads
to a decoupling of the equations and to a block-diagonalization of the mass matrix defining the
kinetic energy. This is achieved by variations of ξ α with zero variations of ṙI , i.e. setting the shape
velocity to zero, which in geometric terms is equivalent to variations in the vertical space of the
principle bundle. This is formalized using a connection, as described next.

(b) The Mechanical Connection and Locked Velocity

Introducing a connection is not as obvious as in case of kinematic constraint. The starting point
is the reduced Lagrangian l(rI , ξ α , ṙI ) = T (rI , ξ α , ṙI ) − V (rI ) with potential energy V (rI ), and
kinetic energy
! !
I α I 1  T T L (r) K (r) ξ
T (r , ξ , ṙ ) = ξ ṙ (5.3)
2 KT (r) S (r) ṙ

expressed in terms of the mass matrix

!
L (r) K (r)
M (r) = (5.4)
KT (r) S (r)
 ∂l
which defines a G-invariant metric on Q. The momentum Πα = ∂ξ α associated to the fiber is thus
14
Π (r, ξ, ṙ) = L (r) ξ + K (r) ṙ. If this relation of ξ and ṙ cannot be integrated to define a relation
of g and r, the momentum is said to be non-holonomic.

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
For a floating multibody system, for example, where ξ α are twist coordinates of the
floating base body, the momentum co-screw Π ∈ g∗ = se∗ (3) comprises the angular and linear
momentum. One can then introduce the locked velocity Ω = (Ω α ) ∈ g such that L (r) Ω = L (r) ξ +
K (r) ṙ. The name stems from the observation that Ω is the base body velocity which generates the
same momentum Π when the system is regard as a rigid body, i.e. when ṙ = 0. In the humanoid
example, this is the velocity the hip would attain, when all joints are locked instantaneously.
Accordingly, L (r) : g → g∗ is called the locked inertia tensor, while S is the inertia related to the
shape coordinates, and K is the cross coupling inertia. If ξ α are coordinates of the base twist in
body-fixed representation, i.e. left-invariant, then Ω ∈ g is usually called the locked body velocity.
This change of coordinates on g is formalized by means of a connection Amech = Adg (g −1 dg +
−1
Adr) on the principal bundle Q. The local connection one-form is defined with (Aα I ) := L K so
that the velocity shift is
Ω α = ξ α + Aα I I
I (r )ṙ . (5.5)

The so-defined connection, is called the mechanical connection [8,9] building upon a concept
discussed in [62]. In contrast to the kinematic connection, it is defined via the momentum. The
connection may be considered to be in Chaplygin form since it is independent of group variables.
Notice that this strictly relies on a local bundle trivialization since Q may not be a trivial principal
bundle. The locked velocity Ω is the vertical part relative to the mechanical connection.

Remark 5.1. An important aspect of the locked velocity is that it cannot be associated to a frame whose
motion is described by h (g, r) ∈ G depending on some g and r, so that Ω̂ = h−1 ḣ. This is an immediate
consequence of the fact that Ω is defined by the non-holonomic momentum Π (r, ξ, ṙ) (see Sec. 5(e)).

Remark 5.2. The principal bundle view on floating-base systems has an interesting connection to gauge
theory. In gauge theory, the symmetry is related to gauge invariance, G is called the ’gauge group’, and
the connection one-form A in (5.5) is the ’gauge potential’ [47]. This was discussed for the falling cat and
similar non-holonomic control systems in [60], and more generally, for ’deformable bodies’ (mechanical
structures that can change their shape) in [63,64], and for Maxwell or Yang-Mills fields in [1]. In gauge
theory, the equivariance condition on the connection describes a gauge transformation from a local gauge
A to a new A0 = Adg (g −1 dg + Adr) [19,65]. In case of mechanical systems, it describes a transformation
from one body-fixed frame to another in which velocities are measured. In [63,64], Aβ
I was called the master
gauge, while its curvature is considered as field strength.

(c) Hamel Coefficients and the Mechanical Connection

Here again, the original definition (2.6) of the Hamel coefficients is employed to derive relations
that are today obtained using the machinery of geometric mechanics.

Lemma 5.1. The Hamel coefficients in bundle coordinates (Ω α , ṙI ) are γab
I
≡ 0, and
α
γβδ = ±cα
βδ (5.6)
α α
γJβ = −γβJ = ±cα δ
βδ AJ (5.7)
∂Aα ∂Aα ∂Aα ∂Aα
α
γIJ = I
− J α
+ γJβ Aβ
I =
I
− J
± cα β λ
βλ AI AJ (5.8)
∂rJ ∂rI ∂rJ ∂rI
where the positive sign of ± applies to left-, and the negative sign to a right-trivialization of the G-bundle.
Proof. Canonical coordinates sα are introduced on G, which are related to the fiber coordinates 15
via ξ α = Aα α β α α α β
β (s ) ṡ and ṡ = Bβ (s ) ξ , respectively (see Sec. 4(b)). Relation (5.5) and its inverse
are then written as

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
Ω α = Aαβ s
α
β
ṡ + Aα J I
I (r )ṙ (5.9)

ṡα = Bβα sa Ω β − Bβα sa Aβ J I

I (r )u (5.10)

With the locked velocity, the quasi-velocities are (ua ) = (uα , uI ) = (Ω α , ṙI ), thus expressed in the
form (2.1), with BJI = δJI and BβI ≡ 0, or in matrix form
  ! ! !     !
Aα β (AαI) ṡ ṡ Bβα − Bβα Aβ I
u= , = u. (5.11)
0 I ṙ ṙ 0 I

The relation (2.6) is separated for the coordinates Ω α and ṙI . The coefficients γβδ α
=
 ∂Aα α
µ ∂Aλ µ λ α
∂sλ
− ∂sµ Bβ Bδ = ±cβδ are again determined by the structure constants of G, and thus

∂Aα ∂Aα
 
α
γβJ = δ
− λ
Bβλ Bγδ AγJ = ±cα δ
βδ AJ .
∂sλ ∂sδ
α
The remaining coefficients γIJ are given in (4.3).
α α
The Hamel coefficients (5.8) are identical to the components of the curvature (1.3), BIJ = γIJ ,
of the mechanical connection in bundle coordinates. They are indeed formally identical to the
curvature coefficients (4.3) of the kinematic connection.

(d) Boltzmann-Hamel Equations as reduced Euler-Lagrange Equations in

Bundle Coordinates
Denote with `(rI , Ω α , ṙI ) := l(rI , ξ α := Ω α − Aα I I
I ṙ , ṙ ) the reduced Lagrangian in terms of the
α
locked velocity Ω . The Hamel equations (2.4) are
d ∂` ∂`  β I β

α
+ β
γαI ṙ + γαλ Ωλ = Qα (5.12)
dt ∂Ω ∂Ω
d ∂` ∂` ∂`  β J β

I
− I + β
γIJ ṙ + γIα Ωα = QI . (5.13)
dt ∂ ṙ ∂r ∂Ω
The motion in G is obtained from the reconstruction equations (4.10), now with ξ α = Ω α −
Aα I
I (r) ṙ defined by the mechanical connection. It is important that the motion is deduced from
ξ , and not from Ω α , as the latter can in general not be attributed to a frame motion (Rem. 5.1
α

and Sec. 5(e)). Equations (5.12,5.13), admit the following geometric interpretation.

Proposition 5.1. The Hamel equations (5.12,5.13) for a left G-invariant Lagrangian `(rI , Ω α , ṙI ) are the
reduced Euler-Lagrange equations (1.1,1.2) in terms of bundle coordinates (Ω α , ṙI ), with the coefficients
α α
EβI = γIβ = cα λ α α α α
βλ AI , cβλ = γβλ , and curvature BIJ = γIJ determined by the Hamel-coefficients.

The equations (1.1,1.2) have been derived in [9], and presented in [33, p. 397], by introducing
the locked velocity (5.5) into the Lagrangian `(rI , Ω α , ṙI ) before taking the Euler-Lagrange
derivative. Their derivation as Hamel equation in terms of the locked velocity has not been
reported in the literature. The first equation (1.1), respectively (5.12), is indeed the Euler-Poincaré
equation (5.1) with ξ α replaced by the velocity Ω α of the locked system, which is why (1.1,1.2) are
β
also called Lagrange-Poincaré equations [45, p. 3395], [3, p. 146]. The terms with γαI in (5.12)
and (5.13) can be regarded as interaction (or coupling) terms. Clearly, as remarked in [4, pp.
912, 913], the equations (5.12,5.13) reduce to the Hamel equations (5.1,5.2) if the coefficients of
connection and curvature vanish, i.e. when expressed in local coordinates (ξ α , ṙI ). However, since
the Hamel formalism applies to any choice of local coordinates, as shown in this paper, it should
not be said that (5.12,5.13) reduce to the Hamel equations when using local coordinates (ξ α , ṙI ), 16
as occasionally stated, e.g. [66, p. 226].
It follows from the definition of the Hamel coefficients that the mechanical connection is flat,

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
α ∂` ∂l
i.e. the curvature BIJ vanishes, if and only if the momentum Πα = ∂Ω α (equivalently Πα = ∂ξ α )

defines a non-integrable relation of ξ α and ṙI .

α
Remark 5.3. A note on the sign convention for the curvature coefficients BIJ is in order. Given a
α α α
connection, the local curvature is usually defined as BIJ = −γIJ with γIJ in (4.3) (see [47] for right
bundles), which agrees with the definition of curvature used in gauge theory [19, p. 247]. This convention
α
is used in [9, p. 157] and [45, p. 3395], and thus BIJ appears with a positive sign in the reduced Euler-
Lagrange equations (1.2). In [4, p. 910] and [18, notice the correction on p. 44], the local curvature is
α α
introduced as BIJ = γIJ , as in this paper, along with the reduced Euler-Lagrange equations (1.1,1.2).
α α
In [3, pp. 117,146], the curvature is derived as BIJ = −γIJ , but is then used with a negative sign in the
α α
reduced Euler-Lagrange equations. Similarly in [66], the curvature is introduced as BIJ = γIJ but used
with positive sign in (1.2). These inconsistencies deserve particular attention when applying equations
(4.9) and (1.1,1.2).

(e) Inertial Decoupling of Poincaré and Lagrange Equations

The mechanical connection allows to intrinsically split variation into the vertical and horizontal
parts4 . As a consequence, the mass matrix M of the equations in terms of the locked velocity, is
diagonal [3, p. 147]. Inserting (5.5) into (5.3) yields the mass matrix
!
L (r) 0
MΩ (r) = (5.14)
0 S (r) − AT (r) L (r) A (r)
 
and the kinetic energy `(rI , Ω α , ṙI ) = 12 Ω T L (r) Ω + 21 ṙT S (r) − AT (r) L (r) A (r) ṙ. Indeed,
∂` α
the momentum Πα = ∂Ω α only depends on the fiber coordinates Ω , so that the equations (5.12)

and (5.13) are inertially decoupled (not coupled on accelerations level). Coupling of the equations
β
is via the velocity terms involving γαI . Inertial decoupling using the locked velocity has been
addressed for modeling of floating base robots [13,67] and space robots in [68,69].
A closely related concept for decoupling the equations is that of the centroidal momentum as
introduced in [70], which is widely used for whole-body control of humanoid robots [15,71,72]
for instance. In this context g = se (3), and V ∈ se (3) is the velocity (twist) of the base body
(using symbol V instead of ξ), and ṙ are the joint velocities. A frame FG is introduced that is
located at the total COM of the system and aligned with the inertial frame F0 . The configuration
of FG relative to the frame Fb attached at the base body is described by gbG ∈ SE (3) = G.
∗
The centroidal momentum is defined as ΠG = AdT gbG Π, with momentum co-screw Π ∈ se (3)
defined in Sec. 5(b). This is also expressed as ΠG = MG VG , where VG is referred to as the
average velocity, and MG = AdT gbG Mbb AdgbG is called the centroidal composite inertia matrix,
and AdT gbG (L, K) the centroidal momentum matrix [14,15]. Comparing this with the definition of
the locked velocity L (r) Vloc = Π shows that VG = Ad−1 gbG Vloc . Moreover, VG = (ωave , ṗcom ),
where ωave = ωloc is called the average angular velocity [70], and ṗcom is the velocity of the total
!
ΘG 0
COM of the system, both expressed in F0 . The important point is that MG = is
0 m̄I
a block diagonal matrix, where m̄ is the total mass, and ΘG is the total inertia tensor w.r.t. to the
total COM expressed in F0 . This would replace the locked inertia L in (5.14) when the EOM are
expressed with VG .
The motivation for using the centroidal momentum ΠG = (ΘG ωave , m̄ṗcom ) is that the linear
and angular momentum are decoupled (in addition to the inertial decoupling of (5.12) and
4
The vertical space is the tangent space ker Tq π to the group orbits, i.e. possible velocities of the base body for locked shape
coordinates. The horizontal space is the space of velocities not producing a locked velocity.
(5.13)), and can be controlled independently. However, there is generally no frame associated 17
to VG (Rem. 5.1) that could serve to represent the system orientation. This would imply that the
−1
motion of this frame is represented by a gbG (g, r) such that V̂G = ġbG gbG . It is clear from the

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
definition of VG (and Vloc ) by means of the momentum that such a frame exists if and only
if the momentum defines integrable relation, i.e. if the curvature vanishes, which is generally
not the case. This seemingly obvious fact was proven in [73]. In order to determine the base
configuration g ∈ SE (3) w.r.t. F0 , the reconstruction equations (4.10), which are now V̂b = g −1 ġ,
must be solved with Vb = Vloc − L−1 Kṙ, as proposed in [71]. Finally it should be remarked that
the centroidal kinematics and dynamics can be expressed in terms of barycentric vectors [74,75].

(f) Example: Satellite with three symmetric reaction wheels

The simplified model of a satellite equipped with three reaction wheels is considered. Fig. 2 shows
a schematic drawing of the principle mechanical setup. In the following, the reaction wheels are
called rotors, for simplicity. The axes of the three rotors are mutually orthogonal, the rotors are
located arbitrarily at the satellite, and are assumed to be symmetric (so that the total COM of
the satellite is constant). The satellite’s main body is modeled as a rigid body. It is assumed that
there are no gravity or other potential forces acting on the satellite, thus the Lagrangian is the
kinetic energy. The motion of the main body is a rigid body motion evolving in a Lie group
G, and represented by g ∈ G. The kinetic energy is invariant w.r.t. G-actions. The rotations of
the three rotors are described by the rotation angles ϕi , i = 1, 2, 3. The configuration space of
the satellite model is thus Q = G × T 3 , with configuration q = (g, ϕ), with ϕ = (ϕ1 , ϕ2 , ϕ3 ). The
¯
latter serve as coordinates on the shape space, rI := ϕI . Different choices for G are used in the
literature. Most of the original formulations in multibody system dynamics use the direct product
group G = SO (3) × R3 , while recent research uses the proper rigid body motion group SE (3)
(mainly triggered by development of Lie group integration methods [76], and geometrically
exact modeling of Cosserat continua [77,78]). The particular choice of symmetry group, but also
whether left- or right trivialization is used, leads to different definition of rigid body velocities
and equations of motion. In the following, notation from multibody dynamics and robotics is
adopted, where V ∈ R6 ∼ = g denotes velocity of a frame (rigid body), and g is either so (3) × R3 or
se (3). A detailed description and numerical results can be found in the supplement [16].

3
b
1 2

3 2

Figure 2. Schematic drawing of a satellite model. The figure shows a self-stabilizing cube reported in [79].
(i) Mixed Representation of Rigid Body Velocity –Symmetry Group G = SO (3) × R3 18
3
The configuration (pose) of the main body is represented as (R, p) ∈ SO (3) × R , where R ∈

rspa.royalsocietypublishing.org Proc R Soc A 0000000

SO (3) and p ∈ R3 describes the rotation and translation of a body-fixed reference frame (RFR)

..........................................................
Fb relative to an inertial frame (IFR) F0 . The group multiplication on the direct product group
is (R1 , p1 ) · (R2 , p2 ) = (R1 R2 , p1 + p2 ). Since rotations and translations are decoupled, this
is clearly not a frame transformation (i.e. a rigid body motion). The corresponding velocity
defined via left-trivialization is V̂b = g −1 ġ = (R−1 Ṙ, ṗ) = (ω̂, ṗ) ∈ g = so (3) × R3 , and in vector
representation Vb = (ω, ṗ) ∈ R6 , where ω ∈ R3 is the angular velocity of the main body relative
to F0 resolved in Fb . This is referred to as the mixed representation of rigid body velocities since
ω is resolved in the body frame, and ṗ in the inertia frame [30,31]. Regarding the dynamics,
decoupling of rotation and translation is valid only if the body-fixed RFR is located at the COM,
which is the main premise when using the direct product group G, since then the angular and
linear momenta are decoupled. Therefore, the velocities of main body and rotors will be measured
at the total COM of the satellite (main body including the rotors), thus p is the position vector of
the total COM resolved in the IFR, and Fb is located at the total COM. Denote with Vi = (ωi , ṗ)
the hybrid velocity of rotor i = 1, 2, 3. W.l.o.g. the RFR is aligned with the rotor axes. Then
ωi = ω + ei ϕ̇i , where ei ∈ R3 is the i-th unit vector (e.g. e1 = (1, 0, 0)), and thus Vi = Vb + Vi ,
with Vi = (ei ϕ̇i , 0).
The momentum of the main body in mixed representation is Π b = Mb Vb ∈ R6 ∼ = g∗ =
∗ 3 i i
so (3) × R , and of the i-th rotor Π = M Vi , with the inertia matrix of the main body and
the i-th rotor, respectively,

! !
b Θb 0 i Θi 0
M = , M = (5.15)
0 mb I 0 mi I

where Θ b and Θ i is the inertia tensor of the main body and the i-th rotor w.r.t. the total COM,
and mb and mi is the mass of the main body and i-th rotor.
The velocity coordinates are (ξ α , ṙI ) = (ω 1 , ω 2 , ω 3 , ṗ1 , ṗ2 , ṗ3 , ϕ̇1 , ϕ̇2 , ϕ̇3 ) = (Vbα , ϕ̇i ) = (Vb , ϕ̇),
with fiber coordinates (ξ α ) = (Vbα ) = (ω 1 , ω 2 , ω 3 , ṗ1 , ṗ2 , ṗ3 ), α = 1, . . . , 6 and (ṙI ) = (ϕ̇1 , ϕ̇2 , ϕ̇3 ),
I = 7, 8, 9. In the following, indexes i, j, k, l = 1, 2, 3 and the notation I¯ = I − 6 are used. The
kinetic energy of the satellite is

3
1 T b 1X T i
T (Vb , ϕ̇) = Vb Π + Vi Π
2 2
i=1
3
1 b α β 1 Xh i i
= Θαβ ω ω + Θjk (ω j + δij ϕ̇i )(ω k + δik ϕ̇i ) + m̄ṗi ṗi (5.16)
2 2
i=1

where m̄ := mb + 3i=1 mi is the total mass of the satellite. The kinetic energy is invariant under
P

left-action of G (due to the body-fixed angular velocity).

Hamel Equations The structure coefficients on the direct product group SO (3) × R3 are cβ
αλ =
εαλβ , for α, β, λ = 1, 2, 3, and cβ
αλ = 0 otherwise. The Euler-Poincaré equations (5.1) are found as

3
d ∂T ∂T j k X j j
+ c ω = Θ̄ij ω̇ j + εikj ω k Θ̄jl ω l + (Θij ϕ̈ + εikl ω k Θlj ϕ̇j )
dt ∂ω i ∂ω j ik
j=1

d ∂T
= m̄p̈j (5.17)
dt ∂ ṗj
where Θ̄ := Θ b + 3i=1 Θ i is the composite inertia tensor of the satellite including main body
P
19
and rotors. They can also be written in vector form, with matrix adω = ω,
e

rspa.royalsocietypublishing.org Proc R Soc A 0000000

3

..........................................................
d ∂T ∂T X i i 
− adT
ω = Θ̄ ω̇ + ω
e Θ̄ω + e i ϕ̇i
θ ϕ̈ + ωθ (5.18)
dt ∂ω ∂ω
i=1
d ∂T
= m̄p̈ (5.19)
dt ∂ ṗ
where θ i := Θ i ei is the i-th column of Θ i .
¯
The Euler-Lagrange equations (5.2) are, noting that T does not depend on ϕi (= rI ),
d ∂T i j i i
= Θij ω̇ + Θii ϕ̈ (5.20)
dt ∂ ϕ̇i
i
where the diagonal element Θii is the moment of inertia about the axis of rotator i.
The above equations are summarized to the set of EOM for the satellite
θ1 θ2 θ3
 
Θ̄ 0 e Θ̄ω + 3i=1 ωθe i ϕ̇i
  P 
ω̇ ω
 0 m̄I 0 0 0 


 1T  p̈ 
 
 0 

 θ 1 1
 0 Θ 11 0 0  
 ϕ̈

 +

 0  = 0.

(5.21)
 2T

2   ϕ̈2  
0

 θ 0 0 Θ22 0    
3T 3
θ 0 0 0 Θ 3 ϕ̈ 0
33

The mass matrix has the form (5.4) with non-zero submatrix K. It is constant due to the
assumption of symmetric rotors and axes aligned with the RFR axes.
The pose of the satellite is obtained by solving
 the kinematic
 reconstruction equations (4.10). To
this end, the equations Ẋ = dexp−1 −1
−X Vb = dexp−x ωb , ṗ are solved for the coordinate vector
∼ so (3) × R3 . Here dexp is the matrix form of the right-trivialized differential
X = (x, p) ∈ R6 = x
of the exp map on SO (3) [55].

Euler-Lagrange Equations on a Trivial Principle Bundle With the partitioning (5.4) of the mass
matrix, the connection coefficients Aα
I are defined by
!
−1 θ̄ 1 θ̄ 2 θ̄ 3
A=L K= . (5.22)
0 0 0
¯
The locked velocity (5.5) is Vloc = Vb + L−1 Kϕ̇, and thus with (5.22), ωloc α
= ω α − Aα I
I ϕ̇ and
α α
ṗloc = ṗ . In terms of the locked velocity, the kinetic energy is
1 b  α I¯

β β J¯

T (Vloc , ϕ̇) = Θαβ ωloc − Aα¯
I ϕ̇ ω loc − A J¯
ϕ̇ (5.23)
2
3
1X i  α α
I¯

β J¯
 m̄
+ Θαβ ωloc + δIα ¯ − AI¯ ϕ̇ ωloc + (δJβ¯ − Aβ
¯)ϕ̇
J
+ ṗα ṗα .
2 2
i=1

A straightforward calculation yields, with block matrices S, K, L deduced from (5.21),

!
∂T Θ̄ 0 ∂T    
= Vloc , = S − AT LA ϕ̇ = S − KT L−1 K ϕ̇ (5.24)
∂Vloc 0 m̄I ∂ ϕ̇

with T T T
1
− θ 1 θ̄ 1 −θ 1 θ̄ 2 −θ 1 θ̄ 3
 
Θ11
S − KT L−1 K = 
T T T
−θ 2 θ̄ 1 2
− θ 2 θ̄ 2 −θ 2 θ̄ 3 (5.25)
 
Θ22 
T T T
3 1 3 2 3 3 3
−θ θ̄ −θ θ̄ Θ33 −θ θ̄
Thus the mass matrix in the EOM in terms of the locked velocity has the block-diagonal form
(5.14). The Hamel coefficients (5.6)-(5.8) are determined by the non-zero structure coefficients
cβ
αλ = εαλβ , for α, β, λ = 1, 2, 3. Since the connection coefficients are constant, the curvature
 α
coefficients (5.8) are BIJ = [AI , AJ ]α . In vector representation, AI is the I-th
¯ column in (5.22), and
20
I¯ J¯
[AI , AJ ] = (θ̄ × θ̄ , 0). It is non-zero due the non-commutativity of vector fields AI w.r.t. the
Lie bracket on g = so (3) × R3 (non-parallel rotor axes). The non-zero BIJ α
(non-flat connection)

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
implies that the momentum does not define an integrable relation of rotor and base motion.

(ii) Body-fixed Representation of Rigid Body Velocity — Symmetry Group G = SE (3)

The semi-direct product group SE (3) = SO (3) n R3 describes proper rigid body motions.
The configuration of the main body is again represented as (R, p) ∈ SE (3), but with group
multiplication (R1 , p1 ) · (R2 , p2 ) = (R1 R2 , p1 + R1 p2 ), which correctly accounts for coupling
of rotations and translations. Thus, the body-fixed RFR can be located arbitrarily. The velocity
(also called twists) of the main body, i.e. of Fb , in body-fixed representation [30,31] is
defined via left-trivialization as V̂b = g −1 ġ = (R−1 Ṙ, R−1 ṗ) = (ω̂, v) ∈ g = se (3), and in vector
representation Vb = (ω, v) ∈ R6 , where now v ∈ R3 is the linear velocity of the main body relative
to F0 resolved in Fb . To simplify the derivation, the velocity of main body and rotors are
expressed in the body-fixed RFR at the main body. The velocity of rotor i = 1, 2, 3 is Vi = (ωi , vi ).
Assuming again that the RFR is aligned with the rotor axes, it holds true that Vi = Vb + Vi .
The momentum of the main body in body-fixed representation is Π b = Mb Vb ∈ R6 ∼ = g∗ =
∗ i i
se (3), and of the i-th rotor Π = M Vi , with the inertia matrix of the main body and of the i-th
rotor w.r.t. an arbitrary RFR
! !
b Θb mb deb i Θi mi d
ei
M = , M = (5.26)
−mb d eb mb I −mi d e i mi I

where Θb and Θi are the inertia tensors of the main body and the i-th rotor w.r.t. the RFR, and db ,
di are the position vectors to the COM w.r.t. the RFR. The total kinetic energy of the satellite is
3 3
1 1X T i 1 1X
T
T (Vb , ϕ̇) = VbT Mb Vb + Vi M Vi = VbT Mb Vb + Vb + Vi (ϕ̇) Mi Vb + Vi (ϕ̇) .


2 2 2 2
i=1 i=1
(5.27)

Hamel Equations In the following, the matrix form of the equations will be presented, for
simplicity. The structure coefficients on the semi-direct product group SE (3) give rise to the
matrix form of the adjoint operator [55,80]
!
ωeb 0
adVb = (5.28)
v
eb ω eb

so that the Lie bracket is adX Y = [X, Y]. The Euler-Poincaré equations (5.1) are
3 3
d ∂T ∂T ˙ − adT LV + X Mi V
 
− adT Mi V
X
Vb = LV̇b + i Vb b i (5.29)
dt ∂Vb ∂Vb
i=1 i=1
P3
with locked mass matrix L = Mb + i
i=1 M . Written explicitly yields the instructive form
3    
e˙ + ωe θ i ϕ̈i + ωθ
e i − mi v
e ai ϕ̇i
X
e Θ̄ω − m̄(v
Θ̄ ω̇ + ω e v)d+ = 0
i=1
  X3  
m̄ v̇ + ωv e˙ + ω
e + (ω e ω)d
e − mi ai ϕ̈i + ωa
e i ϕ̇i = 0 (5.30)
i=1

with Θ̄ := Θ b + 3i=1 Θ i and θ i := Θ i ei as above, and ai := d

P e i ei , where d := (mb db +
P3
i=1 di mi )/m̄ is the position vector of the total COM measured in the RFR.
 The Euler-Lagrange equations (5.2) are found immediately as 21
d ∂T i i
= Θii ϕ̈ + ω̇ T θ i − mi aT
i v̇ (no summation over i). (5.31)
dt ∂ ϕ̇i

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
Clearly, if d = 0, i.e. the RFR Fb is located at the total COM, these equations are equivalent to
those in (5.18-5.20) when modeling the system on G = SO (3) × R3 .
In matrix form, the motion equations are

θ1 θ2 θ3
 
Θ̄ m̄d
  
ω̇ ∗
e
 −m̄d m̄I −m1 a1 −m2 a2 −m3 a3   
  ∗∗ 
v̇ 
 e  
 1T T 1
 1  
 θ −m a Θ 0 0 ϕ̈  +  0  = 0 (5.32)

 1 1 11


 2 T   ϕ̈2   0 
−m2 aT 2
  
 θ 2 0 Θ22 0  
θ3
T
−m aT 0 0 Θ3 ϕ̈3 0
3 3 33
P3 i i P3 i
with ∗ := ω
e Θ̄ω − m̄ωe
e vd + i=1 (ωθ
e − mi v
e ai )ϕ̇ and ∗∗ := m̄ (ωv
e +ω e −
e ωd) i=1 mi ωa
e i ϕ̇ .

Eule-Lagrange Equations on a Trivial Principle Bundle The mass matrix in (5.32) is block-
partitioned, according to (5.4), with
 
1
!
1 2 3
! Θ11 0 0
Θ̄ m̄d
e θ θ θ 2
L= , K= , S= 0 Θ22 0 .
 
−m̄de m̄I −m1 a1 −m2 a2 −m3 a3 3
0 0 Θ33

Therewith, the local connection, defining the locked velocity Vloc = Vb + Aϕ̇ in (5.5), is
 
A = L−1 K = m̄1 m̄2 m̄3 (5.33)
!
ei
with column vectors m̄i := L−1 Mi . Explicit expressions for L, K, A are given in the
0
supplement [16]. The kinetic energy expressed with the locked velocity is
3
1 1X
T
(Vloc − Aϕ̇)T Mb (Vloc − Aϕ̇) + Vloc + Vi (ϕ̇) − Aϕ̇ Mi Vloc + Vi (ϕ̇) − Aϕ̇ .


T (Vloc , ϕ̇) =
2 2
i=1
(5.34)
The partial derivatives in (5.12) and (5.13) are found (replacing Ω α with Vloc
α
and ṙI with ϕ̇i , i = I)
¯
as
3
∂T
Mi Vi = LVloc
X
= L (Vloc − Aϕ̇) + (5.35)
∂Vloc
i=1
∂T
and ∂ ϕ̇as in (5.24). Consequently, the mass matrix becomes block diagonal is in (5.14). The Hamel
coefficients (5.6-5.8) are determined by the structure constants cα βδ on SE (3). Again, the curvature
α
BIJ = [AI , AJ ]α does not vanishing because of the non-commutativity of G = SE (3), where
¯ ¯ ¯
[AI , AJ ] = [m̄I , m̄J ] = adm̄Ī m̄J is the Lie bracket on se (3) in (5.28), i.e. screw product [80]. The
Euler-Lagrange equations (5.12),(5.13) are thus determined explicitly. Finally, the inverse of the
locked mass matrix attains the closed form
!
−1 U −Ude
e −1 .
L = 1 , with U = (Θ̄ + m̄d e d) (5.36)
dU m̄ I − dUd
e e e

The satellite pose is obtained by solving the local kinematic reconstruction equations
Ẋ = dexp−1 6∼
−X Vb for the instantaneous screw coordinate vector X = (x, y) ∈ R = se (3), see
supplement [16]. The matrix form of the dexp map on SE (3) also possesses a closed form [55].

6. Floating-Base Mechanical Systems with Symmetry and

Conserved Momentum
(a) Hamel Equations, Lagrange–d’Alembert–Poincaré equations 22
Conservation laws can be used to introduce a connection. For floating systems with G-invariant

rspa.royalsocietypublishing.org Proc R Soc A 0000000

Lagrangian l(rI , ξ α , ṙI ), the momentum Π ∈ g∗ , in local bundle coordinates, is

..........................................................
∂l
Πα = = Lαβ ξ β + KαJ ṙJ . (6.1)
∂ξ α
Assuming that the initial momentum is zero, the momentum conservation Πα = 0 imposes
non-holonomic dynamic constraints uα = 0, α = 1, . . . , m̄, which are expressed in terms of the
mechanical connection with
uα = ξ α + Aα I I
I (r )ṙ . (6.2)

The Hamel equations are the Lagrange–d’Alembert–Poincaré equations (4.9), now with the
curvature of the mechanical connection in (5.5). The connection encodes dynamic constraints
due to the momentum conservation. If the initial momentum is zero, then the locked velocity
is also zero. Comparing ξ α = uα − Aα I α α α I
I ṙ , obtained from (6.2), with ξ = Ω − AI ṙ , obtained
from (5.5), shows that the equations (4.9) are obtained from the equations (5.13), in terms of the
locked velocity, when Ω α is set to zero and the mechanical connection is used:
d ∂` ∂` ∂` β J
− I + B ṙ = QI (6.3)
dt ∂ ṙI ∂r ∂Ω β IJ
β β
with `(rI , Ω α , ṙI ) := l(rI , ξ α := Ω α − Aα I I
I ṙ , ṙ ), and curvature BIJ = γIJ given by the Hamel
α
coefficients in (5.8), where Ω is set to zero after taking the derivatives. The system dynamics is
thus described in terms of coordinates rI on the base manifold (shape space). The motion in G
is determined as solution of the kinematic reconstruction equations (4.10) with ξ α = −Aα I (r) ṙ
I

defined by the dynamic constraints Πα = 0.

As example, consider the satellite in Sec. 5.(f), with Lagrangian ` equal to the kinetic energy T .
According to (6.3), the Hamel equations in terms of the rotor angles ϕi , i = I¯ are given with
d ∂T ∂T β J¯
− B ϕ̇ = QI (6.4)
dt ∂ ϕ̇I¯ ∂V β IJ
loc

with ∂∂T
ϕ̇i
∂T
, and ∂V α in (5.24) if G = SO (3) × R3 , and with ∂V
∂T
α in (5.35) if G = SE (3). The
loc loc
α
components of the curvature are the Hamel coefficients γIJ in (5.8) given with the structure
constants of the respective symmetry group G.

(b) Geometric Phase and Pseudo-Holonomic Motion

The significance of the mechanical connection on the principal bundle is that it reveals the
geometric phase shift (holonomy) dg = −gAdr, i.e. the motion in the fiber, as a result of the motion
along a closed curve in shape space (Rem. 4.2), which is proportional to the curvature (here
written for left-trivialization). This is due to non-integrable condition imposed by the momentum
conservation (while for constrained systems this is due to non-holonomic kinematic constraints,
Rem. 4.2). Whether a closed path in shape space Vδ̄ leads to a closed path in G is a question arising
in context of motion planning of space robots. Although for non-holonomic systems, this is not
possible globally, there may be trajectories that show such cyclicity. This phenomenon was given
the attribute pseudo-holonomic, and necessary conditions were reported in [81,82] for planar space
robots. This aspect was not treated in the literature for general space robots performing spatial
motions. In view of (4.8), it follows from the mean value theorem that a necessary condition
is the existence of a point r0 ∈ Vδ̄ within the area enclosed by the closed path in shape space
such that the curvature of the mechanical connection vanishes, i.e. B (r0 ) = 0. How this can be
translated into cyclic ’pseudo-holomic’ path planning is topic of current research. As a simple
example, a floating base robot equipped with an arm comprising two revolute joints with parallel
axes is discusses in the supplement [16]. For this space robot, a simple cyclic motion of the two
joints leads to a pseudo-holonomic behavior so that the base motion is also cyclic (zero geometric
phase). That is, along this path the base motion is a function of the arm motion, despite the 23
momentum conservation imposing a non-holonomic constraint.

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
(c) Non-Zero Momentum and the Dynamic Phase
Equations (6.3) apply also when the total momentum is non-zero. The centroidal momentum
0
ΠG = const is the conserved quantity, which is related to its body-fixed representation by
Π 0 (g) = Ad−T 0
gbG ΠG (Sec. 5(e)). The net change of group variables is determined by the extended
reconstruction equations
dg = gL−1 Π 0 dt − g Adr
d (6.5)

that replace equations (4.10). Solving the reconstruction equations for a full cycle along a closed
path in Vδ̄ yields the total phase shift as in (4.8), but now with the additional term gL−1 Π 0 . The
latter delivers the dynamic phase which is intrinsically due to the (initial) momentum, and leads
to a symmetry breaking from G to the symmetry group that preserves the initial momentum. If
(6.5) is regarded as a control problem, this term is the drift vector field. As an example, consider
the satellite in Sec. 5(f) with specific parameters. The rotation of the wheels is prescribed as r (t) =
(π (cos (2πt) − 1) , π sin (2πt) , π/2 sin (4πt)), which is periodic with cycle time 1 s. The geometric
and dynamic parameters, and animations can be found in the supplementary material [16]. First
assume zero total momentum. The motion of the base (i.e. of base frame Fb located at geometric
center of the base body, as shown in Fig. 2) is found from the reconstruction equations (4.11).
Fig. 3a) shows the translation of Fb in the x-y-plane of F0 over 6 s time duration, i.e. for six
cycles of the rotor motion, starting at the origin. Indicated is the position after each cycle, which
corresponds to the translation component of the geometric phase. The translation of Fb is caused
by the rotation about the total COM, which is not the origin of Fb . Fig. 3b) shows the translation
when the initial momentum is not zero. As an example, the momentum is set to Π 0 = Kṙ (0),
which is the momentum injected by the rotors when the base is at rest. This resembles the situation
where a satellite is released with spinning fly-wheels. The base motion is caused by the turning
rotors via the non-holonomic kinematics as well as the dynamics due to the momentum, which
determine the total phase. For completeness, the translation that is generated by the conserved
momentum only when the rotors are rest, is shown as dashed line, which yields the dynamic
phase.

2
y [mm]
y [mm]

4
Dynamic
1 2 phase
Total
phase
0 0

-2
-1 Geometric
phase
-4
-2
-1 0 1 2 3 0 2 4 6 8 10 12 14
a) x [mm] b) x [mm]

Figure 3. a) Translation of base frame Fb projected onto the x-y -plane of F0 (Fig. 2), when the total momentum is zero,
and Fb and F0 initially coincide. Positions after full cycles (with 1 s) of the rotor motion (geometric phase) are indicated.
b) Base motion for non-zero initial momentum (solid line), with positions after a full cycle of rotor motions indicated. Shown
separately (dashed line) is the motion only due to the initial momentum, and the corresponding dynamic phase.
7. Remark on Classical Riemannian Geometry Formulations 24

It should be recalled that geometric approaches to analytical dynamics of discrete mechanical

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
systems have a long history. They were originally developed in the setting of Riemannian
geometry for unconstrained systems with Lagrangian that is quadratic in q̇ a [21,83,84], where the
configuration space Vn is treated as a Riemannian space with metric induced by the Lagrangian.
This was later extended to systems in non-holonomic quasi-velocities and non-holonomically
constrained systems, and Hamel’s equations are viewed as the Lagrange-d’Alembert equations on
a configuration manifold whose tangent space is defined by non-holonomic constraints (which are
in classical literature called non-holonomic tangent bundles). An overview of classical coordinate
formulations can be found in [22,25], and using modern notations of differential geometry in [85].
Only a few publications deal with rheonomic systems and with systems where the Lagrangian is
non-quadratic in q̇ a . Such systems are modeled in the n + 1-dimensional event space Vn × R+ ,
which is treated as a Finsler space. Thus the corresponding model-based control schemes are
developed in event space [86]. Also in this classical setting, the connection and its curvature play a
key roll. For example, consider an unconstrained holonomic system with Lagrangian L (q a , q̇ a ) :=
1 a b
2 gab (q) q̇ q̇ defined by the kinetic energy. The mass matrix defines a Riemannian metric with
coefficients gab on the configuration space Vn . The system dynamics, on the holonomic tangent
bundle T Vn , is governed by the equations
D a
q̇ = Qa (7.1)
dt
where Dξ a = dξ a + Γbc a b c
ξ dq is the absolute differential of a contravariant vector field ξ a , and the
generalized forces Q = g ab Qb . The Christoffel symbols of second kind Γbc
a a
define a natural affine
connection, which is metric and symmetric (Γbc = Γcb ), thus the configuration space Vn is torsion
a a

free. While this is a classical result, there is a beautiful relation for the linearized equations, which
is less known. Denote with (xa ) ∈ Rn small perturbations superposed to the nominal trajectory
q a , so that q a (t) + xa (t) is the perturbed trajectory. The linearized equations along the nominal
trajectory q a are, in covariant form, with the Riemann-Christoffel curvature tensor Rcbda ,

D 2 xb  
gab 2
+ Rcbda q̇ c q̇ d − ∇b Qa xb = Φa (7.2)
dt
the covariant derivative ∇b Qa = ∂Q a
∂q b
− Γbac
Qc , and small generalized forces Φa dual to xa . The
curvature is hence a measure of stability of the perturbed dynamics.

8. Conclusion
The classical Hamel formulation is a generally applicable approach in analytical mechanics for
describing the dynamics of finite-dimensional systems in terms of local coordinates, which can
be extended to continua [87]. Frequently, the coordinate form of equations that can be derived
coordinate-free in the framework of geometric mechanics are referred to as Hamel equations. The
link between these conceptually very different approaches has not been sufficiently addressed,
however. This link was established in this paper, where the key is to identify the Hamel
coefficients as the coefficients appearing in the coordinate form of the reduced Euler-Lagrange
equations, respectively the Lagrange-Poincaré equations. Of particular significance are the local
curvature coefficients that are central in many aspects of control and computational treatment
of mechanical systems whose configuration space is a non-linear manifold or a Lie-group.
Therewith, a clear connection between the equations governing the dynamics on a principle
bundle, defined by a connection originating from certain symmetries, and the original Hamel
formulation is established. In this context the choice of bundle coordinates is crucial. As such the
locked velocity, and the related concept of average velocity, were discussed. The locked velocity
leads to inertial decoupling, which is important for control and computational investigations.
This should motivate further research into their use for deriving formulations with improved
efficiency. A problem that is increasingly receiving attention is that the average velocity cannot be 25
used for kinematic reconstruction [67,71,73,88,89]. This could, for instance, be addressed by means
of holonomy minimizing gauge transformations, i.e. introducing a frame that is not body-fixed

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
nor aligned with the inertia frame. As a geometric aspect of the motion of non-holomic systems, it
was discussed how the geometric phase leads to attitude change of floating systems for instance,
and that there may be pseudo-holonomic motions. Since this is naturally covered by the geometric
approach, it shall motivate treating Hamel’s formalism in a geometric setting. It remains to be
explored how Hamel’s formulation can be extended to the general case when constraints and
Lagrangian possess (possibly complementary) symmetries, as treated in [9,18], where a non-
holonomic connection is introduced generalizing the kinematic and mechanical connection. As a
side-contribution, some differences and inconsistencies of the definition of local curvature found
in the literature were identified, which is crucial when applying equations (1.1,1.2).

A. List of Symbols
n – number of (generalized) coordinates
m̄ – i) number of Pfaffian constraints, ii) dimension of the symmetry group G
δ̄ = n − m̄ – differential (instantaneous) DOF defined by the m̄ Pfaffian constraints
a, b, c, . . . – indices running over all coordinates: a = 1, . . . , n
I, K, L, . . . – indices I = m̄ + 1, . . . , n of i) independent velocity, ii) shape coordinates
(i.e. coordinates of the base manifold of the principle bundle) of a
constrained or unconstrained system
α, β, γ, . . . – indices α = 1, . . . , m̄ of i) constraint equations, ii) dependent velocity
coordinates, iii) canonical coordinates on the symmetry group G
i, j, k, l, . . . – indices i, j, k, l = 1, 2, 3 of Cartesian vectors, e.g. x = (xi ) ∈ R3
q a , q = (q a ) – local coordinates, generalized coordinates of unconstrained system
rI , r = (rI ) – independent (local) coordinates, I = m̄ + 1, . . . , n
sα , s = (sα ) – dependent (local) coordinates, α = 1, . . . , m̄
a α α
γbc , γIJ , γβJ – Hamel coefficients
α
u – i) quasi-velocities, ii) bundle coordinates, iii) Pfaffian constraints
Ωα – local coordinates of the locked velocity
Aα α
I , BIJ – coefficients of the local connection and of the local curvature
εijk – Levi-Civita symbol
δij – Kronecker delta symbol
x
e – skew symmetric matrix x e = (εikj xk ) associated to vector x = (xk ) ∈ R3
x×y – cross product of x, y ∈ R3 , can be written as x ey
η̂ ∈ g – element of Lie algebra g, corresponding to vector η ∈ Rn ∼ =g
cα
βλ – structure coefficients of the Lie algebra g
[X, Y ] – Lie bracket [X, Y ] = cα β λ
βλ X Y of X, Y ∈ g
F0 , Fb – Inertial frame (IFR) F0 , body-fixed frame Fb
SE (3) – special Euclidean group (rigid body motion group) SE (3) = SO (3) n R3
SO (3) – special orthogonal group (rotation group)
Ricci’s summation convention: e.g. BIa uI = I BIa uI , cα β λ
P P P α β λ
βλ X Y = β λ cβλ X Y

References
1. R. Montgomery, J. E. Marsden, and T. S. Ratiu, “Gauged Lie-Poisson structures,” Contemp.
Math, vol. 28, pp. 101–114, 1984.
2. R. W. Montgomery, The bundle picture in mechanics. University of California, Berkeley, 1986.
3. A. M. Bloch, Nonholonomic mechanics and control. Springer, 2003.
4. W.-S. Koon and J. E. Marsden, “Optimal control for holonomic and nonholonomic mechanical
systems with symmetry and Lagrangian reduction,” SIAM Journal on Control and Optimization,
vol. 35, no. 3, pp. 901–929, 1997.
 5. N. Sreenath, “Nonlinear control of planar multibody systems in shape space,” Mathematics of 26
Control, Signals and Systems, vol. 5, no. 4, pp. 343–363, 1992.
6. E. A. Shammas, H. Choset, and A. A. Rizzi, “Towards a unified approach to motion

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
planning for dynamic underactuated mechanical systems with non-holonomic constraints,”
The International Journal of Robotics Research, vol. 26, no. 10, pp. 1075–1124, 2007.
7. A. Guichardet, “On rotation and vibration motions of molecules,” in Annales de l’IHP Physique
théorique, vol. 40, no. 3, 1984, pp. 329–342.
8. J. E. Marsden, Lectures on Mechanics. Cambridge, 1992.
9. J. Marsden and J. Scheurle, “The reduced Euler-Lagrange equations,” Fields Institute Comm.,
vol. 1, pp. 139–164, 1993.
10. G. Hamel, “Über nichtholonome Systeme,” Math. Annalen, vol. 92, no. 1, pp. 33–41, 1924.
11. ——, Theoretische Mechanik: Eine einheitliche Einführung in die gesamte Mechanik. Springer-
Verlag Berlin Heidelberg, 1949.
12. D. V. Zenkov, “On Hamel’s equations,” Theoretical and Applied Mechanics, vol. 43, no. 2, pp.
191–220, 2016.
13. H. Mishra, M. De Stefano, A. M. Giordano, R. Lampariello, and C. Ott, “A geometric controller
for fully-actuated robotic capture of a tumbling target,” in 2020 American Control Conference
(ACC). IEEE, 2020, pp. 2150–2157.
14. D. E. Orin and A. Goswami, “Centroidal momentum matrix of a humanoid robot: Structure
and properties,” in 2008 IEEE/RSJ Int. Conf. on Intel. Rob. and Sys., 2008, pp. 653–659.
15. D. E. Orin, A. Goswami, and S.-H. Lee, “Centroidal dynamics of a humanoid robot,”
Autonomous robots, vol. 35, no. 2, pp. 161–176, 2013.
16. A. Müller, “Supplement to "Hamel’s Equations and Geometric Mechanics of Constrained and
Floating Multibody and Space Systems",” Figshare, DOI: 10.6084/m9.figshare.c.6607449.
17. D. D. Holm, Geometric mechanics-Part I: Dynamics and symmetry. World Scientific, 2011.
18. A. M. Bloch, P. Krishnaprasad, J. E. Marsden, and R. M. Murray, “Nonholonomic mechanical
systems with symmetry,” Archive for Rational Mech. and Anal., vol. 136, no. 1, pp. 21–99, 1996.
19. J. C. Baez and J. P. Muniain, Gauge fields, knots and gravity. World Scientific, 1994, vol. 4.
20. T. Frankel, The Geometry of Physics. Cambridge, 2004.
21. E. T. Whittaker, A treatise on the analytical dynamics of particles and rigid bodies (4th ed.).
Cambridge, 1988.
22. J. Papastavridis, Analytical Mechanics: A Comprehensive Treatise on the Dynamics of Constrained
Systems for Engineers, Physicists, and Mathematicians. Oxford University Press, 2002.
23. A. Sommerfeld, Vorlesungen über theoretische Physik, Band 1: Mechanik,. Akademische
Verlagsgesellschaft, Leipzig, 1948.
24. G. Hamel, “Über die virtuellen Verschiebungen in der Mechanik,” Math. Ann., vol. LIX, pp.
416–434, 1904.
25. P. Maisser, “Differential-geometric methods in multibody dynamics,” Nonlinear Analysis:
Theory, Methods & Applications, vol. 30, no. 8, pp. 5127–5133, 1997.
26. J. M. Maruskin and A. M. Bloch, “The boltzmann-hamel equations for optimal control,” in
2007 46th IEEE Conference on Decision and Control. IEEE, 2007, pp. 554–559.
27. L. Boltzmann, “Über die Form der Lagrange’schen Gleichungen für nicht holonome
generalisierte Coordinaten,” Sitzungsberichte der Mathematisch-Naturwissenchaftlichen Klasse der
Kaiserlichen Akademie der Wissenchaften in Wien, vol. 111, pp. 1603–1614, 18. December 1902.
28. ——, Vorlesungen über die Principe der Mechanik: II Teil. Verlag von Johann Ambrosius Barth,
Leipzig, 1904.
29. G. Hamel, “Die Lagrange-Eulersche Gleichungen der Mechanik,” Z. Math. Phys., vol. 50, no. 1,
pp. 1–57, 1904.
30. R. Murray, Z. Li, and S. Sastry, A Mathematical Introduction to Robotic Manipulation. CRC Press,
1994.
31. A. Müller, “Screw and Lie group theory in multibody kinematics,” Multibody System Dynamics,
vol. 43, no. 1, pp. 37–70, 2018.
32. ——, “On the Hamel Coefficients and the Boltzmann–Hamel Equations for the Rigid Body,”
Journal of Nonlinear Science, vol. 31, no. 2, pp. 1–39, 2021.
33. J. E. Marsden and T. S. Ratiu, “Introduction to mechanics and symmetry,” Physics Today,
vol. 48, no. 12, p. 65, 1995.
34. T. G. Ionescu, P. Antonescu, I. Biro, G. Bögelsack, and A. K. Breteler, “Terminology for the
mechanism and machine science,” Mech. Mach. Theory, vol. 38, pp. 767–901, 2003.
35. A. Müller and D. Zlatanov, Singular Configurations of Mechanisms and Manipulators. Springer, 27
2019.
36. C. Ehresmann, “Les connexions infinitésimales dans un espace fibré différentiable,” in

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
Colloque de topologie, Bruxelles, vol. 29, 1950, pp. 55–75.
37. A. M. Bloch, J. E. Marsden, and D. V. Zenkov, “Nonholonomic dynamics,” Notices of the AMS,
vol. 52, no. 3, pp. 324–333, 2005.
38. R. L. Hatton and H. Choset, “Geometric motion planning: The local connection, stokes
theorem, and the importance of coordinate choice,” The International Journal of Robotics
Research, vol. 30, no. 8, pp. 988–1014, 2011.
39. S. A. Chaplygin, “On a motion of a heavy body of revolution on a horizontal plane,” Regular
and Chaotic Dynamics [English translation], vol. 7, no. 2, pp. 119–130, 2002.
40. ——, “On some generalization of the area theorem with applications to the problem of rolling
balls (matematicheskii sbornik, vol. 20, 1897, pp. (in russian)),” English translation: Regular and
Chaotic Dynamics, vol. 17, no. 2, pp. 199–217, 2012.
41. P. Voronets, “Equations of motion for nonholonomic systems,” Matematicheskii Sbornik, vol. 22,
no. 4, pp. 659–686., 1901.
42. P. Woronetz, “Über die Bewegung eines starren Körpers, der ohne Gleitung auf einer
beliebigen Fläche rollt,” Mathematische Annalen, vol. 70, no. 3, pp. 410–453, 1911.
43. S. Soltakhanov, M. Yushkov, and S. Zegzhda, Mechanics of non-holonomic systems – A New Class
of control systems. Springer, 2009.
44. E. A. Shammas, H. Choset, and A. A. Rizzi, “Towards a unified approach to motion
planning for dynamic underactuated mechanical systems with non-holonomic constraints,”
The International Journal of Robotics Research, vol. 26, no. 10, pp. 1075–1124, 2007.
45. J. E. Marsden, T. S. Ratiu, and J. Scheurle, “Reduction theory and the lagrange–routh
equations,” Journal of mathematical physics, vol. 41, no. 6, pp. 3379–3429, 2000.
46. J. McCauley, Classical Mechanics. Cambridge, 1997.
47. Y. Choquet-Bruhat, C. DeWitt-Morette, and M. Dillard-Bleick, Analysis, Manifolds and Physics
- Part I. North-Holland, 1996.
48. S. D. Kelly and R. M. Murray, “Geometric phases and robotic locomotion,” Journal of Robotic
Systems, vol. 12, no. 6, pp. 417–431, 1995.
49. J. Ostrowski and J. Burdick, “Geometric perspectives on the mechanics and control of robotic
locomotion,” pp. 536–547, 1996.
50. G. Darboux, “Leçons sur la théorie généles applications géométriques du calcul infinitesimal,”
Gautiers-Villars, Paris, vol. 4, 1887.
51. D. Condurache, “Poisson-Darboux problems’s extended in dual Lie algebra,” in AAS/AIAA
Astrodynamics Specialist Conference, Stevenson, WA, USA, 2017.
52. H. Munthe-Kaas, “Runge-Kutta methods on Lie groups,” BIT Numerical Mathematics, vol. 38,
no. 1, pp. 92–111, 1998.
53. ——, “High order Runge-Kutta methods on manifolds,” Applied Numerical Mathematics,
vol. 29, no. 1, pp. 115–127, 1999.
54. A. Iserles, H. Z. Munthe-Kaas, S. P. Nørsett, and A. Zanna, “Lie-group methods,” Acta
Numerica, vol. 9, p. 215–365, 2000.
55. A. Müller, “Review of the exponential and Cayley map on SE (3) as relevant for Lie group
integration of the generalized Poisson equation and flexible multibody systems,” Proceedings
of the Royal Society A, vol. 477, no. 2253, 2021.
56. D. McDonald, “How does a man twist in the air?” New Scientist, vol. 10, no. 237, pp. 501–503,
1961.
57. T. Kane and M. Scher, “A dynamical explanation of the falling cat phenomenon,” International
journal of solids and structures, vol. 5, no. 7, pp. 663–670, 1969.
58. H. Essén, “The cat landing on its feet revisited or angular momentum conservation and
torque-free rotations of non-rigid mechanical systems,” American Journal of Physics, vol. 49,
no. 8, pp. 756–758, 1981.
59. M. Enos, “On an optimal control problem on SO(3)xSO(3) and the falling cat,” American
Mathematical Society (Ed.): Dynamics and Control of Mechanical Systems, 1993.
60. R. Montgomery, “Gauge Theory of the Falling Cat,” Fields Inst. Commun, vol. 1, pp. 193–218,
1993.
61. D. Gerritsen and M. Kuipers, “On the angular motion of a freely falling human or animal
body,” Journal of Engineering Mathematics, vol. 13, no. 4, pp. 347–353, 1979.
62. S. Smale, “Topology and mechanics. ii,” Inventiones mathematicae, vol. 11, no. 1, pp. 45–64, 1970. 28
63. A. Shapere and F. Wilczek, “Self-propulsion at low reynolds number,” Physical Review Letters,
vol. 58, no. 20, p. 2051, 1987.

rspa.royalsocietypublishing.org Proc R Soc A 0000000

..........................................................
64. ——, “Gauge kinematics of deformable bodies,” American Journal of Physics, vol. 57, no. 6, pp.
514–518, 1989.
65. C. J. Isham, Modern differential geometry for physicists, 2nd ed. World Scientific Publishing
Company, 1999, vol. 61.
66. H. Cendra, J. E. Marsden, and T. S. Ratiu, “Geometric mechanics, Lagrangian reduction, and
nonholonomic systems,” in Math. unlimited—2001 and beyond. Springer, 2001, pp. 221–273.
67. G. Garofalo, B. Henze, J. Englsberger, and C. Ott, “On the inertially decoupled structure of the
floating base robot dynamics,” IFAC-PapersOnLine, vol. 48, no. 1, pp. 322–327, 2015.
68. A. M. Giordano, D. Calzolari, and A. Albu-Schäffer, “Workspace fixation for free-floating
space robot operations,” in 2018 IEEE International Conference on Robotics and Automation
(ICRA). IEEE, 2018, pp. 889–896.
69. A. M. Giordano, C. Ott, and A. Albu-Schäffer, “Coordinated control of spacecraft’s attitude
and end-effector for space robots,” IEEE Rob. and Aut. Let., vol. 4, no. 2, pp. 2108–2115, 2019.
70. H. Essén, “Average angular velocity,” European journal of physics, vol. 14, no. 5, p. 201, 1993.
71. G. Nava, L. Fiorio, S. Traversaro, and D. Pucci, “Position and attitude control of an
underactuated flying humanoid robot,” in 2018 IEEE-RAS 18th International Conference on
Humanoid Robots (Humanoids). IEEE, 2018, pp. 1–9.
72. S.-H. Lee and A. Goswami, “Reaction mass pendulum (rmp): An explicit model for centroidal
angular momentum of humanoid robots,” in Proceedings 2007 IEEE International Conference on
Robotics and Automation. IEEE, 2007, pp. 4667–4672.
73. A. Saccon, S. Traversaro, F. Nori, and H. Nijmeijer, “On centroidal dynamics and integrability
of average angular velocity,” IEEE Rob. Automat. Letters, vol. 2, no. 2, pp. 943–950, 2017.
74. S. Dubowsky and E. Papadopoulos, “The kinematics, dynamics, and control of free-flying and
free-floating space robotic systems,” IEEE Tran. Rob. Automat., vol. 9, no. 5, pp. 531–543, 1993.
75. E. G. Papadopoulos, “Nonholonomic behavior in free-floating space manipulators and its
utilization,” in Nonholonomic Motion Planning. Springer, 1993, pp. 423–445.
76. E. Celledoni and B. Owren, “Lie group methods for rigid body dynamics and time integration
on manifolds,” Computer Methods in Appl. Mech. and Eng., vol. 192, no. 3, pp. 421 – 438, 2003.
77. M. Borri and C. Bottasso, “An intrinsic beam model based on a helicoidal approximation-Part
I: Formulation,” Int. J. for Num. Methods in Eng., vol. 37, no. 13, pp. 2267–2289, 1994.
78. V. Sonneville, A. Cardona, and O. Brüls, “Geometrically exact beam finite element formulated
on the special Euclidean group SE(3),” Computer Methods in Applied Mechanics and Engineering,
vol. 268, pp. 451–474, 2014.
79. H. Gattringer, A. Reiter, C. Stöger, M. Jörgl, P. Hörmandinger, and A. Müller, “Dynamical
modeling and swing-up control of a self-balancing cube,” in International Conference on Robotics
in Alpe-Adria Danube Region. Springer, 2016, pp. 144–151.
80. J. M. Selig, Geometric fundamentals of robotics. Springer Science & Business Media, 2004.
81. R. Mukherjee and M. Zurowski, “Pseudo-holonomic behavior of planar space robots,” in 1994
IEEE Int. Conf. on Rob. and Automat. IEEE, 1994, pp. 2405–2410.
82. R. Mukherjee, “Pseudoholomorphic behavior of planar space robots,” Journal of Guidance,
Control, and Dynamics, vol. 19, no. 1, pp. 251–253, 1996.
83. J. L. Synge and A. Schild, Tensor calculus. University of Toronto Press, 1949.
84. A. I. Lurie, Analytical Mechanics (in Russian). State Publishing House, Physical And
Mathematical Literature, Moscow, 1961; Published as English translation: A.I. Lurie,
Analytical Mechanics, Springer, 2002.
85. F. Bullo and A. D. Lewis, Geometric control of mechanical systems: modeling, analysis, and design
for simple mechanical control systems. Springer, 2004, vol. 49.
86. U. Jungnickel, G. Kielau, P. Maisser, and A. Müller, “A passivity-based control of euler-
lagrange systemswith a non-quadratic lagrangian,” ZAMM-Journal of Applied Mathematics and
Mechanics, pp. 982–992, 2008.
87. D. Shi, D. V. Zenkov, and A. M. Bloch, “Hamel’s formalism for classical field theories,” Journal
of Nonlinear Science, vol. 30, no. 4, pp. 1307–1353, 2020.
88. J. I. Mulero-Martínez, “A new factorization of the coriolis/centripetal matrix,” Robotica,
vol. 27, no. 5, pp. 689–700, 2009.
89. W. Du, Z. Wang, E. Moullet, and F. Benamar, “Meaningful centroidal frame orientation of
multi-body floating locomotion systems,” in IEEE Int. Conf. Rob. Automat, 2021, pp. 3061–3067.