Symbolic Computation of Dynamics on

Smooth Manifolds

Brian Bittner and Koushil Sreenath

Dept. of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, USA.

Abstract. Computing dynamical equations of motion for systems that

evolve on complex nonlinear manifolds in a coordinate-free manner is
challenging. Current methods of deriving these dynamical models is only
through cumbersome hand computations, requiring expert knowledge of
the properties of the configuration manifold. Here, we present a symbolic
toolbox that captures the dynamic properties of the configuration man-
ifold, and procedurally generates the dynamical equations of motion for
a great variety of systems that evolve on manifolds. Many automation
techniques exist to compute equations of motion once the configuration
manifold is parametrized in terms of local coordinates, however these
methods produce equations of motion that are not globally valid and
contain singularities. On the other hand, coordinate-free methods that
explicitly employ variations on manifolds result in compact, singularity-
free, and globally-valid equations of motion. Traditional symbolic tools
are incapable of automating these symbolic computations, as they are
predominantly based on scalar symbolic variables. Our approach uses
Scala, a functional programming language, to capture scalar, vector, and
matrix symbolic variables, as well as the associated mathematical rules
and identities that define them. We present our algorithm, along with its
performance, for computing the symbolic equations of motion for several
systems whose dynamics evolve on manifolds such as R, R3 , S 2 , SO(3),
and their product spaces.

1 Introduction

Computing dynamics of systems directly on nonlinear manifolds involves labo-

rious manual computation and expert knowledge of the properties of the con-
figuration space. The motivation for such computation lies in the end result
of obtaining a compact, globally valid, and singularity free dynamical model.
These traits enable the use of geometric controllers to obtain almost global sta-
bility properties. While such dynamical models for certain popular, modestly
complex mechanical systems are readily available (through hand computations),
there exists no procedural or automatic way to compute the symbolic dynam-
ics on manifolds. The primary reason is that we do not have a way to capture
non-scalar symbolic variables and perform non-scalar symbolic arithmetic. The
contribution of this paper addresses this gap.
2

Fig. 1: The spherical pendulum, quadrotor with suspended load, and three-link
walker systems that evolve on complex nonlinear manifolds. The dynamics for
each are symbolically computed using the presented framework, see codebase
at https://gitlab.com/HybridRobotics/dynamics_on_manifolds.

The paper can be summarized by the following: For systems that evolve
on manifolds (see Figure 1), traditional ways to obtain dynamical models in-
volve taking a local parametrization of the configuration space, such as Euler
angles, and then solving for the dynamics. However, obtaining the dynamics
on manifolds, without local parametrization, results in a dynamical model that
is condensed to a compact expression, is globally valid, and is in a singularity
free form. But, there exist no tools that can automate solving for the symbolic
dynamics on manifolds. We provide a symbolic algebraic framework to capture
geometric axioms and identities that govern the scalar, vector, and matrix
elements of dynamical systems. An algorithm is shown to automate computation
of the dynamics of simple mechanical systems on cartesian products of Cn , Rn ,
S 2 , and SO(3), as well as including computation of dynamics of a set of complex
robotic systems.
The rest of the introduction places our proposed algorithm amongst prior
work involving dynamics solvers and computer algebra systems, followed by an
outline of the paper. With respect to dynamics solvers, Euler-Angle parametriza-
tion has long been the standard for system development. Introducing these lo-
cal parametrizations of manifolds enables the use of well-established analytical
methods and corresponding software algorithms for deriving the equations of
motion. For instance, the Newton-Euler method is implemented in Neweul [16]
and SD/FAST [14]. The Recursive Newton-Euler Algorithm has been used to
efficiently compute dynamics of serial chain manipulators, [7,9]. The Lagrangian
method has been implemented on various multibody systems [22]. Additionally,
the Piogram Method [5], the Composite Rigid Body Algorithm [11], and the
Articulated Body Algorithm [11], all provide efficient algorithmic approaches for
solving the dynamical equations of long kinematic chains. Commercial systems
such as ADAMS [24], SimMechanics, MapleSim, LMS Dads [27], etc., not only
solve the equations of motion, but also offer efficient simulations. Symoro [15]
allows for symbolic modeling and simulation for high dimensional kinematic
tree-structures and closed kinematic chains. Although there are several efficient
methods and corresponding software tools that automate the computation of
 3

equations of motion, every single one of them is based on local parametrizations

that employ scalar symbolic variables.
Automating symbolic computation on nonlinear manifolds requires treating
vectors and matrices as single-entity symbolic variables with operations defined
directly on them, which existing computing tools do not support. We achieve this
by creating symbolic variables that represent either scalars, vectors, or matrices,
with additional properties such as constants, unit vectors, symmetric matrices,
etc. Moreover, using a functional programming approach, we enable a program-
matic understanding of mathematical axioms and identities between these vari-
ables.
We looked at the fields best reputed computer algebra systems to encap-
sulate our dynamics framework. Commonly used Computer Algebra Systems
include Mathematica [3] and Matlabs Symbolic Math Toolbox [2], (see [1] for
full list of CASs). These only support scalar symbolic variables (vector and ma-
trix symbolic variables exist only indirectly through scalars.) ATLAS 2 [10],
an extension of Mathematica, enables high quality visualizing of n-dimensional
manifolds, but is not applicable for our computation since it also requires local
frame parametrizations. Our solution is to use Scala, a functional programming
language with a vast set of object oriented programming capabilities [25]. This
enables the design of strong expression classes and an intuitive functional pro-
gramming environment for capturing mathematical identities. This coupled with
existence of an extendable ScalaMathDSL [13] package made Scala preferable to
the frameworks SymPy [30] and Axiom [8]. We extend prior work on scalar ex-
pressions in ScalaMathDSL [13] to handle vector and matrix classes, resulting
in a powerful framework to implement the dynamics algorithm. Note that our
modeling tool does not provide any type of numerical computation.
The rest of the paper is structured as follows: Section II, covers the math-
ematical background. Section III details the dynamics framework and walks
through the algorithm to automate computation of the symbolic dynamics. Sec-
tion IV tabulates results of applying the dynamics solver on a set of mechanical
systems. Finally Sections V and VI provide a discussion and concluding remarks.

2 Mathematical Background

Many robotic systems evolve on nonlinear manifolds. However, we typically com-

pute the dynamics of these systems through local parametrizations, such as Euler
angles. Robotic systems are hindered by local parametrizations that induce sin-
gularities, model complexity, and unwanted features such as the unwinding phe-
nomena [4] during control. The Lagrange-dAlembert principle (see [20], [21], [18]
for more details.), detailed in Section 2.1, allows the computation of dynamics of
systems evolving on complex, nonlinear manifolds to arrive at singularity-free,
globally valid, and compact equations of motion.
4

2.1 Lagrange dAlambert Principle for Computing Dynamics

Let the degrees of freedom for a dynamical system exist on the configuration
manifold Q. Then, the Lagrangian of the system, L : T Q R is defined as L =
T U, where T : T Q R and U : Q R are the kinetic and potential energies
respectively. Here T Q is the tangent bundle of the configuration space. The
virtual work W : Q R captures energy input to the system via actuation. The
dynamical equations of motion can be computed by minimization of the action
integral S through the Least Action Principle [12], where the action integral is
defined as: Z t Z t
S= L dt + Wdt. (1)
0 0
The minimization is done by setting the variation of the action integral to zero,
i.e., S = 0.

2.2 Computation on S 2 Manifold

To concretely illustrate the procedure, we carry out the following computational
steps to derive the equations of motion of the spherical pendulum on S 2 [19].
(To observe the computation on SO(3) the reader is directed to [6].)
Step 1: Supply the Lagrangian (L) and virtual work (W), providing the com-
ponents of the action integral (S) for the spherical pendulum, as given below,
1 2
L= ml (q q) mgl(q e3 ), W = q ,
2
Z t Z t
1 2
S= ml (q q) mgl(q e3 ) dt + q dt.
0 2 0
Here unit vector q S 2 R3 has the property q q = 1 and kinematic constraint
q q = 0. The constant scalars (m, g, l) R correspond to the mass of the
pendulum, magnitude of acceleration due to gravity, and length of the pendulum
respectively. Constant e3 R3 provides the direction of the gravitational field,
while R3 defines the torque via virtual work at the base of the pendulum.
Generalized vector R3 is orthogonal to the evolution of q such that the
variation of q is given as, q = q.
Step 2: Take the variation of the action integral and set to zero:
Z t Z t
S = ml2 ( q q) mgl(q e3 ) dt + q dt = 0.
0 0

Step 3: Replace variations q and q with the generalized vector :

Z t Z t
ml2 (( q + q) q) mgl( q) e3 ) dt + ( q) dt = 0.
0 0

Step 4: Group equations with respect to and and apply simplifications (e.g.:
q q = 0):
Z t
(ml2 q q) + (ml2 q q mglq e3 + q ) dt = 0.
0
 5

Step 5: Perform integration by parts on :

Z t
(ml2 (q q + q q)) + (mglq e3 + q ) dt = 0.
0

Step 6: Collect for and simplify (e.g.: q q = 0):

Z t
(ml2 q q mglq e3 + q ) dt = 0.
0
Step 7: These equations of motion can be extracted to describe the dynamics:
q = q,
2
q (ml q + mgle3 ) = q ,
where the first identity is a kinematic identity (derived from qs action on S 2 )
with being the angular velocity, and the second identity obtained from Step
6 by setting the term that is dotted with as zero. This variation-based com-
putation returns a set of compact, globally-valid equations of motion that are
singularity-free. Having discussed the mathematical background, we will next
describe the computational framework of our dynamics solver that automates
this method.

3 Computational Design of Dynamics Solver

This section will provide a description of the data structures used to represent
mathematical objects and cover the algorithmic techniques that operate on them.
Critical computational steps in the algorithm are covered, followed by a user-
level code snippet that implements the symbolic computation of dynamics on a
spherical pendulum.

3.1 Class Hierarchies for Mathematical Expressions

In Section II we used symbolic scalars (m, l, g) and symbolic vectors (q, q,
e3 ) to compute the dynamics for the spherical pendulum. These two expression
types along with symbolic matrices make up the base classes used to implement
the dynamics solver. The class hierarchy consists of each base class (ScalarExp,
VectorExp, MatrixExp) and its children, which are comprised of a set of case
classes. As we will see, case classes enable pattern matching to encode math-
ematical rules (e.g.: the chain rule for differentiation, scalar and vector triple
product rules, etc.) Each case class can inherit traits to possess additional prop-
erties (e.g.: vectors having unit magnitude or matrices having skew symmetric
properties.) The unit vector trait allows dynamics to be computed on the S 2
configuration space. The matrix traits constant, skew, and symmetric enable
distinction of skew, skew symmetric, and rotation matrices. This enables direct
and intelligent computation of dynamics on SO(3). Each case class serves as
a unique mathematical entity in the symbolic evaluator, see Figure 2 for the
class hierarchy. Next we will discuss how these case classes interact to create a
meaningful data structure for symbolic computation.
6

Fig. 2: All mathematical expressions inherit from a parent class corresponding

to its identity as a scalar, vector, or matrix. Class traits are inherited to add
additional properties to symbolic expressions.
3.2 Abstract Syntax Trees

Our mathematical expressions are stored as abstract syntax trees (ASTs). Math-
ematical operations, such as addition and multiplication, are represented as par-
ent nodes which contain a specified number of leaves appropriate for their respec-
tive functionality. Leaves on the other hand represent symbolic scalars, vectors,
and matrices. Mathematical algorithms are then written as functional programs
with the ASTs serving as the input and output data structures. Figure 3 provides
an AST of the Lagrangian for the spherical pendulum. As detailed next, ASTs
are coupled with pattern matching functions to encode mathematical rules.

3.3 Embedding Mathematical Axioms

Writing the framework in Scala, a functional and object oriented programming

language, we are able to succinctly embed important mathematical axioms gov-
erning the properties of the configuration manifolds. Because the elements of an
AST are each designated a case class with inherited mathematical traits, this
structure can be quickly parsed to enforce mathematical axioms and identities,
which are used to simplify the dynamics and perform calculus operations.

3.4 Mathematical Operations

Each mathematical step in the computational pipeline, is achieved through spe-

cialized pattern matching functions. These functions perform a mathematical
evaluation or restructuring of the AST. Pattern matching via the extensive class
hierarchy for elements of manifold dynamics enables brief, yet powerful functions
that are intuitive to write and understand.
To illustrate this, we will walk through the various functions implemented to
execute each step of the variation-based method. Accompanied by illustrations
of tree manipulation for a spherical pendulum, we discuss the procedural nature
 7

of each step of the algorithm (steps enumerated 1-7). We will talk about trees of
n nodes containing p parents (operands) and e literals (mathematical elements).
Tree Conditioning: For the steps below, the tree conditioning is defined as
a full expansion of the expression, (i.e. a (b + c) is converted to a b + a c).
Isolating the dynamical terms enable succinct, procedural code to be written for
tree operations. Expanding the tree requires traversing the tree (covered next)
and invokes 3 parent operations per expansion (a b + a c). A maximum of
e 1 expansions can take place during traversal. Table 1 provides a condensed
layout of the computational cost for the algorithm.
Tree Traversal: Each algorithm requires traversal of the tree, from the top
down. The general run time for this case is O(n). If we balance the tree and
parallelize the operations (allowed by the structured commutativity above), we
could theoretically assert an O(log(n)) traversal time. The algorithm enables fast
computation of the dynamics (see Results Tables 2 and 3), with most systems
requiring computation time on the order of milliseconds.

Fig. 3: Step 1: Abstract Syntax Tree representation of the Lagrangian of a

spherical pendulum, given by L = 21 ml2 (q q) mgl(q e3 ). ASTs are the data
structure used to represent all symbolic expressions.
Step 1 requires initialization of the Lagrangian AST. This involves
supplying the information of the mechanical configuration and energy of the
system, resulting in the input AST shown in Figure 3. The Lagrangian AST of
e elements is constructed in e 1 operations.
Step 2 takes the variation of the Lagrangian. To respect space con-
straints, we display variation of only the kinetic energy component of the La-
grangian AST in Figure 4a. Taking the variation of an element changes its string
identifier and possibly its case class. For example, q loses qs trait as a unit vec-
tor. This function also applies the chain rule, (u v) = u v + u v. The chain
rule invokes 3 parent operations and can be applied (e 1) times. The variation
function is run on each element once. Taking the time derivative has equivalent
structure to this algorithm, but generates elements of different properties.
Step 3 shown in Figure 4b, applies constraints through substitu-
tion. The tree is searched for a node containing the expression candidate for
8

replacement. In this case, when q is found, it is replaced with q + q.

Substitution allows for the variations and kinematic constraints of the system
to be enforced, along with simplifying expressions such as q q = 0. Each node
is checked once. The tree is then reconditioned to fully expanded form.
Step 4 collects an expression with respect to a vector or scalar.
Figure 4c illustrates collecting for the generalized vector . For each isolated
energy term, each element is checked for a match to the collected expression. A
collected scalar is set as the leftmost multiplied element, allowed by commuta-
tivity. A collected vector is placed on the left hand side of the dot product by
vector algebraic axioms such as the scalar triple product. Each collection requires
2 parent operations. Examples collecting for a include b (a c) = a (b c) for
scalars and b (a c) = a (b c) for vectors.
Step 5 performs integration by parts on the dot product containing the
collected generalized vectors from the previous step. In Figure 4d, the mathe-
matical operation is enforced by negating and integrating while differentiating
the right hand side of the dot product. The result leaves the expression entirely
in terms of generalized vector . Since integration by parts requires two expres-
sions, the negation, integration, and differentiation involved can operate on up to
e 1 elements. The tree is reconditioned again. This allows for the expression
to be collected with respect to which is done in Step 6. This allows for
the right hand side of the dot product to be set to zero, resulting in the dynam-
ical model obtained in Step 7, where the dynamics can be extracted line
by line, one group of terms set to zero for each generalized vector. One parent
operation is required to remove a term from the tree. This can occur e times.

Parent Literal
Process Traversal
Operations Operations
Step 1: - e1 e
Expand: n 3(e 1) -
Step 2: n 3(e 1) e
Step 3: n - e
Expand: n 3(e 1) -
Step 4: n 2(e 1) -
Step 5: n 2(e 1) 3(e 1)
Expand: n 3(e 1) -
Step 6: n 2(e 1) -
Step 7: n e -
Run Time: O(n) O(e) O(e)

Table 1: This table presents the run time of critical steps in the computation.

Linear run time convergence enables broad inspection of high degree of free-
dom dynamics, the laborious computational nature of which made model ex-
traction previously intractable. The coupled simplicity and effectiveness of the
algorithm is enabled by computing on coordinate free configuration spaces. There
is no trigonometric book-keeping, which grows rapidly for systems described on
S 1 and SO(2).
 9

Fig. 4: Illustration of various functions required for implementing the symbolic

dynamics solver. Shown for each function are the input and output
mathematical expressions, the corresponding input and output ASTs (with
changes highlighted in red), and the corresponding code-snippets that invoke
the tree manipulation.
10

3.5 Example with Spherical Pendulum

Listing 1 illustrates the user code for deriving the dynamics of a spherical pen-
dulum. The user provides the Lagrangian and specifies the actuated variables
of the system. These are input to the computeDynamics function (which im-
plements the algorithm detailed in Section 3.4 to derive the dynamics for the
system. Specification of configuration variables and manifold based constraints
is not required, since the constraint information is encapsulated within mathe-
matical type declaration of each variable. This simple example code will enable
users to effortlessly set up problems for solving the dynamics of new robotic
systems. Constraint manifolds are duduced from configuration variable types,
i.e. q S 2 has kinematic constraint q = q. The codebase is available at
https://gitlab.com/HybridRobotics/dynamics_on_manifolds.
o b j e c t SphericalPendulum {
d e f main ( ) {
// d e f i n e c o n s t a n t scalars
v a l g = Cons ( g ) // g r a v i t a t i o n a l c o n s t a n t
v a l m = Cons ( m ) // p o i n t mass
v a l l = Cons ( l ) // l e n g t h a t which p o i n t mass hangs

// d e f i n e v e c t o r s
v a l e3 = CVec ( e3 ) // o r i e n t a t i o n o f g r a v i t y
v a l q = UVec ( q ) // p o i n t mass a c t s on S2
v a l u = Vec ( u ) // v i r t u a l work done on system

// s e t c o n f i g u r a t i o n v a r i a b l e s
// ( s c a l a r s , v e c t o r s , m a t r i c e s )
v a l c o n f i g V a r s = Tuple3 ( L i s t ( ) , L i s t ( q ) , L i s t ( ) )

// define lagrangian
val KE = Num( 0 . 5 ) m l l ( d i f f V ( q ) dot d i f f V ( q ) )
val PE = m g l ( q dot e3 )
val L = KE PE

// s p e c i f y i n f i n i t e s i m a l v i r t u a l work o f system
v a l infWork = Dot ( del taV ( q ) , u )

var eoms = computeDynamics (L , W, c o n f i g V a r s ) }

Listing 1: User code to compute dynamics for a spherical pendulum.

4 Results of Computing Symbolic Models

Having presented an overview of our computational framework for the symbolic

dynamics solver, we next test this across many robotic systems with dynamics
evolving on several different manifolds. Table 1 provides results on simple pen-
dulum systems with dynamics on manifolds S 2 and SO(3). Table 2, on the other
 11

hand, introduces real robotic systems evolving on highly complex manifolds. The
user supplies the Lagrangian, Virtual Work, and configuration variables. The ta-
ble lists the runtime and equations produced for each mechanical system. Next,
we briefly describe the systems tabulated in Table 1, for which our proposed
symbolic algorithm computes the dynamics on manifolds.
4.1 Simple Mechanical Systems
A Spherical Pendulum [6] is the first system considered, its parametrization
is described in section 2.2.
Configuration Manifold: S 2
Kinematic Constraints: q q = 0
Variations: q = q, q = q + q, R3
Energy Terms:
1
T = ml2 (q q), U = mgl(q e3 ).
2
The 3D Pendulum [28] provides a system which acts on the SO(3) manifold.
The pendulum is represented as a rigid body of mass m and symmetric inertia
matrix J R33 . It is pivoted at the point from which vector R3 extends
toward the center of mass. The orientation of the 3D pendulum is specified by
rotation matrix R SO(3) with , M R3 being the angular velocity and
moment input of the 3D pendulum respectively. The generalized vector R3
is utilized to describe the variation of R on SO(3).
Configuration Manifold: SO(3)
Kinematic Constraints: R = R
Variations: R = R, = + , R3
Energy Terms:
1
T = J, U = mgR e3 .
2
A Double Spherical Pendulum [20] is an extension of the singular Spher-
ical Pendulum. A second pendulum with point mass m2 is pivoted from the end
of the first pendulum containing point mass m1 . Each pendulum has a respec-
tive length li from its pivot to point mass. The unit vector qi S 2 points along
each li with its origin at the respective pivots. The torque at the base of each
pendulum is given by i R3 .
Configuration Manifold: S 2 S 2
Kinematic Constraints: qi qi = 0
Variations: qi = i qi , q i = i qi + i qi , i R3
Energy Terms:
1 1
T = (m1 + m2 )l12 q1 q1 + m1 l1 l2 q1 q2 + m2 l22 q2 q2
2 2
U = (m1 + m2 )gl1 q1 e3 + m2 gl2 q2 e3

Similarly, the Double 3D Pendulum [17] is an extension of the Singular

3D Pendulum. The rigid body of mass m2 is pivoted from the point that vector
R3 points to. The rigid bodies orientations at point masses mi are specified
12

Config. Time to Equations of Motion Produced by

System
Manifold DOF Compute our Symbolic Dynamics Solver

q = q
S2 2 0.019 sec.
q (ml2 q + mgle3 ) = q

Spherical
Pendulum [6]

R = R
SO(3) 3 0.015 sec.
J + J = mg RT e3 + M

3D Pendulum [28]

q1 = 1 q1
q1 ((m1 + m2 )l12 q1 + m2 l1 l2 q2 +
(m1 + m2 )gl1 e3 ) = q1 1
S2 S2 4 0.032 sec.
q2 = 2 q2
q2 (m2 l1 l2 q1 + m2 l22 q2 + m2 gl2 e3 ) =
q2 2

Double Spherical
Pendulum [19]

(J1 m1 2 m2 22 )1 +
m2 1 R1T R2 2 2 + 1 (J1 m1 2
m2 21 )1 + m1 g R1T e3 +
SO(3)
6 0.033 sec. m2 g l1 R1T e3 + m2 1 R1T R2 22 = M1
SO(3)
(J2 m2 22 )2 m2 2 R2T R1 1 1 +
2 (J2 m2 22 )2 + m2 g 2 R2T e3 +
m2 2 R2T R1 12 1 = M2
Double 3D
Pendulum [17]
Table 2: The equations of motion were computed on these simple systems using
the proposed symbolic computation tool with the kinematic constraints,
variations, and Lagrangian (T-U) provided for each system.
 13

by rotation matrices Ri .
Configuration Manifold: SO(3) SO(3)
Kinematic Constraints: R = R
Variations: Ri = Ri i , i = i i + i , i R3
Energy Terms:
1 1 1
T = 1 J1 1 + 2 J2 2 + m2 x2 x2
2 2 2
U = m1 gR1 1 e3 + m2 gR2 2 e3

4.2 Robotic Systems

Having seen the algorithm work for simple mechanical systems, we next consider
robotic systems with more complex configuration spaces, shown in Table 3.
The Three Link Walker [23] represents an idealized mechanical configu-
ration for bipedal walkers. Masses m, mH , and mT , are dynamically actuated
by torques 2 and 3 . Unit vectors q1 , q2 , and q3 specify the motion of the three
links. The Stance Dynamics are as shown below. The Flight Dynamics exist
on S 2 S 2 S 2 R3 , and account for the walkers evolution in Euclidean Space.
Configuration Manifold: S 2 S 2 S 2
Kinematic Constraints: qi qi = 0
Variations: qi = i qi , q i = i qi + i qi
Energy Terms:
5m mH mT 2 m m mT
T =( + + )l q1 q1 + l2 ( q1 + q2 ) q2 + (lLq1 + L2 q2 ) q3
8 2 2 2 8 2
3m m
U =( gl + mH l + mT gl)q1 e3 glq2 e3 + mT gLq3 e3
2 2
The Quadrotor with Tethered Load [29] and Reaction Mass Pendu-
lum [26] can be referred to at the respective sources for information on config-
uration variables and composition of the Lagrangian.
Manifolds Cn and Rn are directly computed by the same steps shown in Sec-
tion III, with the exception of Step 3, which introduces constraint manifolds for
S 2 and SO(3). This step is not required for variables that act on Cn and Rn .
The computational results for each system were validated by a published deriva-
tion and hand calculation. These results establish the validity and efficiency of
computing dynamics on complex manifolds using the symbolic evaluator.

5 Discussion

It should be noted that manually computing and validating the dynamics for
several of the systems in Table 3 would have taken a novice one to several days.
Our proposed method of automating the computation reduces this to fractions
of seconds. This will enable a broader study of novel robotic systems whose
dynamics evolve on complex manifolds, systems that were previously not broadly
14

Config. Time to Equations of Motion Produced by

System
Manifold DOF Compute our Symbolic Dynamics Solver

q1 = 1 q1
q1 (( 54 m+mH +mT )l2 q1 + 21 ml2 q2 +
mT lLq3 + ( 32 m + mH + mT )gle3 ) = 0
q2 = 2 q2
S2 S2
6 0.032 sec. q2 ( 21 ml2 q1 + 14 ml2 q2 + 12 mgle3 =
S2
q2 2
q3 = 3 q3
3-Link Walker q3 (mT lLq1 + mT L2 q3 + mT gLe3 =
Stance q3 3
Dynamics [23]

xL = vL
(mQ + mL )VL + mQ l(q q)q + (mQ +
mL )ge3 = f Re3 + p (p f Re3 )
SO(3)
8 0.024 sec. q = q
S 2 R3
q (ml2 q + mgle3 ) = q T
R = R
Quadrotor with J + J = M
Suspended
Load [29]
q1 = 1 q1
q1 (( 54 m + mh + mt )l2 q1 +
1
2
ml2 q2 + mt lLq3 + ( 74 m + mh +
2mt )lx + ( 32 m + mh + mt )lge3 ) = 0
q2 = 2 q2
q2 ( 21 ml2 q1 + 14 ml2 q2 + 34 mlx +
S2 S2 1
10 0.064 sec. mlge3 ) = q2 2
S 2 R3 2
q3 = 3 q3
q3 (mt lLq1 + mt L2 q3 + 2mt Lx +
3 Link Walker mt Lge3 ) = q3 3
Flight ( 47 lm + lmh + 2lmt)q1 + ( 34 ml +
1
Dynamics [23] 2
mt L)q2 + 32 mt Lq3 + ( 25 m + mh +
4mt )x + (2m + 1mh + 2mt)ge3 = 0

T
m = mL e3 2 L + mgeT3 RL
T
e3
RL = RL L
JL L =
C L JL L + 2me3 2 L +
T T
SO(3) 11 0.073 sec. mge3 RL e3 + L RL RP D
SO(3) S RP = RP P

JP PP = P JP P +
4 mp si si ei 2 P + D
2mP si = 2mP si PT ei 2 P + Ui
Reaction Mass
Pendulum [26]
Table 3: The equations of motion were computed on these robotic systems
using the proposed symbolic computation tool with the kinematic constraints,
variations, and Lagrangian (T-U) provided for each system.
 15

approachable by the robotics community. The compact models will also bring
new insight into the dynamics of existing robotic systems, enabling novel control
designs for achieving highly dynamical maneuvers. A limitation of this method
is that it cannot evaluate systems with nonholonomic constraints. Dynamics can
only be computed on Cartesian products of Cn , Rn , S 2 , and SO(3).

6 Conclusion
We have presented an algorithm which automates the computation of dynami-
cal equations for systems evolving on manifolds. By utilizing pattern matching
within the Scala framework, we are able to capture the geometric axioms and
identities of scalar, vector, and matrix elements of a dynamical system. Using
the framework we implement a generalized algorithm that works across a wide
variety of manifolds. The time efficient computation allows for the software to
provide near-instantaneous output of dynamical equations. The dynamics gener-
ated are compact, globally-valid, and free of singularities, as they are described
directly on the configuration manifold. This tool is released publicly and will
enable broader inspection of systems that act on complex, nonlinear manifolds.

7 Acknowledgements
B. Bittner would like to thank Carnegie Mellons URO, who supported him with
a Summer Undergraduate Research Fellowship. K. Sreenath would like to thank
R. Ravindran for discussions that led to the use of Scala. This work is supported
by NSF grants IIS-1464337 and CMMI-1538869.

References
1. List of computer algebra systems. http://en.wikipedia.org/wiki/List of computer
algebra systems.
2. Symbolic Math Toolbox. Mathworks Inc, 1993.
3. Mathematica Version 10.0. Wolfram Research Inc., 2014.
4. S. P. Bhat and D. S. Bernstein, A topological obstruction to continuous global
stabilization of rotational motion and the unwinding phenomenon, Systems &
Control Letters, vol. 39, no. 1, pp. 6370, 2000.
5. P.-y. Cheng, C.-i. Weng, and C.-k. Chen, Symbolic derivation of dynamic equa-
tions of motion for robot manipulators using piogram symbolic method, IEEE
Journal of Robotics and Automation, vol. 4, no. 6, pp. 599609, 1988.
6. L. Consolini and M. Tosques, On the exact tracking of the spherical inverted
pendulum via an homotopy method, Systems & Control Letters, vol. 58, no. 1,
pp. 16, 2009.
7. P. I. Corke, An automated symbolic and numeric procedure for manipulator rigid-
body dynamic significance analysis and simplification, in IEEE Conf. on Robotics
and Automation, 1996, pp. 10181023.
8. T. Daly, D. V. Chudnovsky, and G. V. Chudnovsky, Axiom The 30 Year Horizon.
The Axiom Foundation, 2007.
16

9. E. Dean-leon, S. Nair, and A. Knoll, User friendly matlab-toolbox for symbolic

robot dynamic modeling used for control design, in IEEE Conf. on Robotics and
Biomimetics, 2012, pp. 21812188.
10. I. DigiArea, Atlas 2 for Mathematica, 2012. [Online]. Available: http:
//www.digi-area.com/Mathematica/atlas/guide/Atlas.php
11. R. Featherstone and D. Orin, Robot dynamics: equations and algorithms, in
IEEE Conf. on Robotics and Automation, 2000, pp. 826834.
12. R. P. Feynman, The principle of least action in quantum mechanics, Ph.D. dis-
sertation, Princeton University Princeton, New Jersey, 1942.
13. T. Flaherty, Scala Math DSL, https://github.com/axiom6/ScalaMathDSL,
2013.
14. M. G. Hollars, D. E. Rosenthal, and M. A. Sherman, SD/FAST users manual.
Symbolic Dynamics Inc, 1991.
15. W. Khalil, F. Bennis, C. Chevallereau, and J. Kleinfinger, Symoro: A software
package for the symbolic modelling of robots, in Proc. of the 20th ISIR, 1989.
16. E. Kreuzer and W. Schiehlen, Neweul software for the generation of symbolical
equations of motion, in Multibody Systems Handbook. Springer, 1990.
17. T. Lee, Computational geometric mechanics and control of rigid bodies, Ph.D.
dissertation, University of Michigan, Ann Arbor, 2008.
18. T. Lee, M. Leok, and N. H. McClamroch, Lagrangian mechanics and variational
integrators on two-spheres, International Journal for Numerical Methods in En-
gineering, vol. 79, no. 9, pp. 11471174, 2009.
19. , Lagrangian mechanics and variational integrators on two-spheres, J. for
Numerical Methods in Engineering, vol. 79, pp. 11471174, 2009.
20. , Stable manifolds of saddle equilibria for pendulum dynamics on S2 and
SO(3), in IEEE Conf. on Decision and Control, 2011, pp. 39153921.
21. , Dynamics and control of a chain pendulum on a cart, in IEEE Conf. on
Decision and Control, 2012, pp. 25022508.
22. R. Lot and M. D. A. Lio, A symbolic approach for automatic generation of the
equations of motion of multibody systems, Multibody System Dynamics, vol. 12,
pp. 147172, 2004.
23. P. X. Miranda, L. Hera, A. S. Shiriaev, B. Freidovich, S. Member, U. Mettin, and
S. V. Gusev, Stable walking gaits for a three-link planar biped robot with one
actuator, IEEE Trans. on Robotics, vol. 29, no. 3, pp. 589601, 2013.
24. G. Nonlinearity, Adams 2014, pp. 15, 2014.
25. M. Odersky and Al., An Overview of the Scala Programming Language, EPFL,
Lausanne, Switzerland, Tech. Rep. IC/2004/64, 2004.
26. A. K. Sanyal and A. Goswami, Dynamics and balance control of the reaction mass
pendulum: A three-dimensional multibody pendulum with variable body inertia,
J. Dyn. Sys., Meas., Control, vol. 136, no. 2, p. 021002, Nov. 2013.
27. D. Services, LMS Virtual . Lab The Unified Environment for Functional Perfor-
mance Engineering, pp. 116, 2007.
28. J. Shen, A. K. Sanyal, N. A. Chaturvedi, D. Bernstein, and H. McClamroch, Dy-
namics and control of a 3d pendulum, in IEEE Conf. on Decision and Control,
2004, pp. 323328.
29. K. Sreenath, T. Lee, and V. Kumar, Geometric control and differential flatness
of a quadrotor UAV with a cable-suspended load, in IEEE Conf. on Decision and
Control, 2013, pp. 22692274.
30. SymPy Development Team, SymPy: Python library for symbolic mathematics,
2014. [Online]. Available: http://www.sympy.org