©

2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any
current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new
collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other
works.

Please cite this paper as:

@INPROCEEDINGS{Scherzinger2019Inverse,
author={S. {Scherzinger} and A. {Roennau} and R. {Dillmann}},
booktitle={19th International Conference on Advanced Robotics (ICAR)},
title={Inverse Kinematics with Forward Dynamics Solvers for Sampled Motion Tracking},
year={2019},
volume={},
number={},
pages={681-687},
arXiv:1908.06252v2 [cs.RO] 12 May 2023

keywords={},
doi={10.1109/ICAR46387.2019.8981554},
ISSN={},
month={Dec},}
 Inverse Kinematics with Forward Dynamics Solvers for Sampled
Motion Tracking
Stefan Scherzinger1 , Arne Roennau1 and Rüdiger Dillmann2

Abstract— Tracking Cartesian motion with end effectors is sparse to serve as direct joint control commands and cause
a fundamental task in robot control. For motion that is not jumps in the robot’s actuators. A remedy is interpolating
known in advance, the solvers must find fast solutions to the in task space, leading again to the problem of decoupled
inverse kinematics (IK) problem for discretely sampled target
poses. On joint control level, however, the robot’s actuators solutions in joint space.
operate in a continuous domain, requiring smooth transitions In this paper we describe an IK approach that inherently
between individual states. In this work, we present a boost achieves both goals for real-time target following. Our ap-
to the well-known Jacobian transpose method to address proach enhances the well-known Jacobian transpose method
this goal, using the mass matrix of a virtually conditioned with a simple but effective component – a conditioned,
twin of the manipulator. Results on the UR10 show superior
convergence and quality of our dynamics-based solver against virtual mass matrix – leading to intuitive solutions of the
the plain Jacobian method. Our algorithm is straightforward IK solver and to smooth intermediate states between sparse
to implement as a controller, using common robotics libraries. targets. To support the vastly growing Robot Operating
System (ROS)4 [10], we accompany the paper with a ROS-
I. INTRODUCTION control [11] controller implementation as power-on-and-go
In robotics, solutions to the Inverse Kinematics (IK) solution for the robotics community5 .
problem are fundamental for manipulator control. Articulated
mechanisms are powered in their joints, and hence require II. P ROBLEM STATEMENT AND RELATED WORK
control algorithms to perform continuous mappings from The forward mapping of the joint positions vector q to
task space to joint space. Being such a core element in Cartesian space is given by
robot control, IK has received significant attention in the last
x = g (q) (1)
decades and has lead to a wide range of different methods.
See e.g. [1] for an overview. Among the fastest approaches such that q = g −1 (x) would represent a close form solu-
are closed form solutions, which are analytically derived tion to the inverse problem, which is to be approximated
solutions, specifically tailored for each robot kinematics. numerically. The velocity vector q̇ maps with
IKFast3 with an example application in [2]. However, closed
form solutions are sensitive to impossible-to-reach targets. ẋ = J (q) q̇ (2)
Numerical approaches provide trade-offs between accuracy to end effector velocity ẋ, using the manipulator Jacobian
and stability, usually leveraging the manipulator Jacobian: J . Furthermore, the common relation
E.g. its transpose [3], [4], [5], its inverse with singularity
stabilizing measures, such as Damped Least Squares (DLS) τ = J T (q) f (3)
[6], [7], [8] or Selectively Damped Least Squares (SDLS) holds for a static end effector force vector f and torques τ
[9]. More recent improvements have shown excellent success in the joints of the robot manipulator. In further notations
rates for general IK solving [8]. While suitable for searching we will drop the joint configuration dependency of J (q) for
vast regions of the solution space in the context of motion the sake of brevity. Using J T to solve (1) for q has been
planning, finding individual joint configurations is not always proposed independently by Wolovich et al [3] and Balestrino
sufficient: et al [4].
Target motion is often sampled into discrete target poses, In [4] the authors describe a dynamical systems according
such as in virtual servoing or end effector teleoperation. to
Solving the IK problem for each target individually leads q̇ = J T Q(xd − g(q)) + w (4)
to valid, but decoupled solutions in joint space with no
guaranteed feasibility of execution. On the other hand, for for coordinate transformations from task- to joint space,
low-frequency sampled targets, individual solutions are too with Q denoting an arbitrary positive definite matrix and
w (Q, J , ) representing an adaptation term, whose exact
1 Stefan Scherzinger and Arne Roennau are with FZI Research
expression we do not repeat here. The authors show asymp-
Center for Information Technology, Haid-und-Neu-Str. 10-14,
76131 Karlsruhe, Germany [email protected] totic stability around the Cartesian desired pose xd with
[email protected]
2 Rüdiger Dillmann is with IAR Institute for Anthropomatics and 3 http://openrave.org/docs/0.8.2/openravepy/ikfast/
4 www.ros.org
Robotics, HIS Humanoids and Intelligence Systems Lab, KIT Karl-
sruhe Institute of Technology, Adenauerring 2, 76131 Karlsruhe, Germany 5 https://github.com/fzi-forschungszentrum-
[email protected] informatik/cartesian_controllers
 = xd − g(q), and validate their approach on a three
DOF mechanism. Although proving effective in converging,
(4) does not offer an intuitive interpretation. Furthermore,
the authors state that velocity tracking is only matched in
the mean, requiring additional filters to deliver real angular
Fig. 1: Closed loop control scheme of our IK solver. The control
velocities for robot control. plant uses simplified forward dynamics computation of a virtually-
In [5] Pechev uses a similar technique with Feedback In- conditioned twin of the real robot.
verse Kinematics (FIK) from a control perspective, deriving

−1 d
q̇ = J T Q J J T Q + I ẋ (5)
respectively. Rearranging for q̈ and using (3) yields the
as system dynamics. In this case Q denotes a dynamic, non- manipulator’s acceleration due to external forces f acting
diagonal matrix that involves an off-line state space optimiza- on its end effector.
tion subject to manipulator dynamics and task requirements.
Solving (5) in a feedback loop as a filter avoids matrix q̈ = H −1 J T f − H −1 C(q, q̇) − H −1 G(q) (8)
inversions, and performs well in singularity experiments Again, we dropped the dependency of q in the notation
in comparison to the DLS method from [6], albeit only for brevity. Solving (8) for q is considered with forward
presented on a three DOF planar manipulator. The Jacobian dynamics, i.e. simulating the mechanism’s motion through
transpose is also commonly used in Cartesian control, see time if loads f are applied. This is commonly achieved
e.g. [12], using the real plants dynamics in these cases. Here, with numerical integration methods and is a vital part in
however, we separate the computation of the IK problem physics engines. The right-hand side of (8) is a superposi-
from the underlying manipulator, which we see as a black tion of acceleration terms in joint space. While taking all
box to be controlled in an open loop fashion with desired terms into consideration is crucial for simulating physically
configurations q d (t) exclusively. correct motion, our approach for solving IK problems has
In [3] the authors derive a solution to the IK problem by a different focus and motivates two simplifications: If we
proposing the usage of J T in form of a simpler dynamical drop acceleration due to gravity (G = 0) and consider
system instantaneous motion only for each cycle (i.e. q̇ = 0), which
q̈ = KJ T xd − g(q)


(6) equals setting H −1 C(q, q̇) = 0, our system still accelerates
They provide a Lyapunov stability analysis and show that in the direction pointed to by f , but that acceleration is not
the system is asymptotically stable in the error dynamics, biased by velocity-dependent non-linearities, nor the effect
using an arbitrary positive definite matrix K. However, they of gravity. To be concise with this simplification, we do not
leave a possible conditioning of K aside and remain with accumulate velocity during time integration in our algorithm
the general proof. In the remainder of the paper we will refer in section IV. Implications of this approach are discussed in
to this approach as the Jacobian transpose method. section VI. Expelling both acceleration terms from (8), we
In this work, we investigate the benefits of choosing K obtain the simplified system
from (6) not to be constant, but instead to be the dynamically
q̈ = H −1 J T f , (9)
changing inertia matrix of a virtually conditioned twin of the
robot manipulator. Like the work in [5] we show the impor- which relates f to instantaneous joint acceleration q̈. Turning
tance of matrices with off-diagonal elements, but introduce the search for IK solutions into a control problem, (9) is the
a more intuitive interpretation, using manipulator dynamics. control plant we chose at the core of our solver.
A primary contribution of this paper is to propose forward We use the distance error  = xd − g(q) as the error
dynamics solvers in the context of J T -based IK solving. between target and current end effector pose. Its entries are
This general idea bases on previous findings [13]. Here, we T
present an enhanced algorithm with improved convergence  = [x , y , z , rx , ry , rz ] (10)
with physically plausible intermediate solutions for motion
with the translational error components and rotational error
control.
components of the according Rodrigues vector.
III. FORWARD DYNAMICS IK SOLVER Relating this distance error vector to a virtual Cartesian
A. Simulation of Robot Motion end effector input
Approximating the manipulator as a set of articulated, f = Kp  + Kd ˙ (11)
rigid bodies, the governing equations of motion obtain the
following form in matrix notation with the positive definite, diagonal gain matrices Kp and
Kd , we obtain
τ = H(q)q̈ + C(q, q̇) + G(q) (7)
q̈ = H −1 J T (Kp  + Kd )
˙ (12)
in which H denotes the mechanism’s positive definite in-
ertia matrix and C and G represent vectors with separated as our forward dynamics motivated controller. Fig. 1 shows
Coriolis an centrifugal terms and gravitational components the closed loop control scheme for solving IK.
 Note the similarity to (6): Setting Kp = I and Kd = 0
yields
q̈ = H −1 J T xd − g(q) .


(13)
However, H −1 generates physically plausible joint accel-
erations q̈ that really reflect the mechanism reaction to
the Cartesian error vector for instantaneous motion. This
means that meeting kinematic constraints, the mechanism
accelerates in Cartesian space with ẍ in the direction as
pointed to by f . In the experiments section we show that this
indeed causes faster and more goal-directed convergence.
B. Homogenization Methods
A goal of this paper is to provide insight on the condition-
ing of the mechanism such that H(q) possesses beneficial
behavior throughout the joint space. Note that unlike the
kinematics of the system, which we can not change, we are
free to give the simulated manipulator any dynamic behavior Fig. 2: Dynamics-conditioned twin of an exemplary mechanism.
The end effector comprises all of the mechanism’s mass m and
we wish. In this context we propose to think of the solver rotational inertia I, which is illustrated with the oversized sphere.
concept from Fig. 1 as using a virtually conditioned twin Links and joints are considered massless, such that their configu-
of the real mechanism for which we wish to solve IK. The ration has a vanishing influence on the overall dynamics.
basic idea behind this approach has been presented in our
previous work [13]. In this paper, we present more detailed
aspects behind this concept and provide an empirical proof IV. IMPLEMENTATION
of this earlier proposition.
The time derivative of (2) gives We implemented the closed loop scheme from Fig. 1 as a
ROS controller in C++. The pseudo code for the IK-solving
ẍ = J˙ q̇ + J q̈ . (14) part is listed in Algorithm 1.
We again consider instantaneous accelerations while the
mechanism is still at rest and set J˙ q̇ = 0. Together with Algorithm 1 IK with forward dynamics
(9) we obtain
1: procedure S OLVER(xd , q0 , ∆t, N )
ẍ = J H −1 J T f , (15)
2: 0 = 0
which describes the Cartesian instantaneous acceleration due 3: for i = 1 to N do
to the controlled distance error f . With setting M −1 = 4: i = xd − g(qi−1 )
J H −1 J T we obtain the more concise notation 5: ˙ i = (i − i−1 )/∆t
ẍ = M −1 f . (16) 6: fi = Kp i + Kd ˙ i
7: q̈i = H −1 (qi−1 )J T (qi−1 )fi
Our intention now is to achieve 8: q̇i = 0.5 q̈i ∆t . q̈i−1 ≡ 0
M −1 (q) to be an almost con- 9: qi = qi−1 + 0.5 q̇i ∆t . q̇i−1 ≡ 0
stant, diagonal matrix for all 10: end for
possible joint configurations q, 11: q d = qN
as illustrated on the right. So 12: return q d
that the gain matrices Kp and 13: end procedure
Kd generate uniform system ac-
celerations ẍ, and our control loop from Fig. 1 has equal
convergence throughout the Cartesian workspace. Note that the computation of H −1 and J T is not part
To achieve this behavior, we follow a mechanically- of the algorithm. For this task we used the Kinematics
motivated assumption: The error-correcting f acts directly on and Dynamics Library (KDL)1 , which is common place
the mechanism’s end effector. If this end effector comprises in ROS. Note that the number of steps N can be chosen
all of the mechanism’s mass and inertia, as illustrated in freely to give the solver (virtual) time for any given xd .
Fig. 2a, then joint configurations will have less effects on M . In combination with the gain matrices Kp and Kd this is
The overall center of mass remains unchanged. This effect is a partially redundant means to tweak the solver to range
depicted in Fig. 2b where f experiences the same rotational from a one-shot IK-solver to an interpolating controller to
inertia M , although having shifted most of the mechanism’s smooth noisy targets xd . We discuss this behavior in the
structure away from the rotatory axis in one case. In the experiment on interpolation performance for low frequency
experiments section we sample a massive number of joint sampled targets.
configurations and show empirically that this approach in-
deed achieves a good homogenization of M −1 = J H −1 J T . 1 http://wiki.ros.org/orocos_kinematics_dynamics
 V. EXPERIMENTS AND RESULTS
In our experiments, we chose the Universal Robot UR10
as manipulator (Fig. 3), which is ubiquitous both in indus-
try and academia, and represents a well-known platform
for many possible users. We want to emphasize, however,
that our proposed IK solver has been used successfully
(a)
on various manipulators. Since a central theme of this
work is to use a dynamics motivated approach for IK 3.2

solving instead of a pure kinematics one, we compare our 2.4

Average μ
method to the Jacobian transpose method where appropriate.
1.6

0.8
A. Controller Homogenization 0.0
In this experiment we evaluated 1.6

our homogenization method and

Variance σ 2
1.2
compared it both to the Jacobian
0.8
transpose method from (6) and a
reference dynamics model. For all 0.4

mechanism, we used the robot’s 0.0

kinematics as provided by the ROS Fig. 3: UR10 1.2
package2 . For the Jacobian trans-
Std. deviation σ
0.9
pose method , we set the gain matrix
0.6
K to identity, which is a reasonable choice without further
knowledge of the system. 0.3

Fig. 4a illustrates the qualitative distribution of mass and

Pure kinematics Naive dynamics Our method
inertia for the experiment. We choose the following values
for our proposed model (Fig. 4a right): (b)

m = 1 kg Fig. 4: Experiment A. Analysis of three mapping approaches from

  f → ẍ for the Universal Robots UR10. To obtain the plots,
1 0 0 (17) we sampled 100.000 random configurations of q uniformly with
I = 0 1 0 kgm2 qi ∈ [−π, π]. For each sample, we computed the mapping matrix
0 0 1 as indicated in (a). The sub figures (b) show mean, variance and
standard deviation (less is better) for the individual matrix entries
which constitute the last link of the manipulator. For numer- for the entirety of samples.
ical stability, we set all other links to m0 = 10−3 m and
I0 = 10−6 I respectively (instead of to zero). The reference
model (Fig. 4a middle) had 1/6 of m and I equally attached target represents a step in Cartesian space, as illustrated
to its links. in Fig. 6 for experiment B. A common use case is low-
Fig. 4b shows the results of the analysis. Note that all mean frequency sample targets. Since the speed of convergence in
matrices are in fact diagonal, as was expected for arbitrary both systems (6) and (12) can be tweaked with the gains
joint space sampling. However, the variances indicate a par- K and Kp , Kd respectively, we chose the following set of
tially strong configuration dependency for Jacobian transpose parameters to make the solvers comparable:
method (pure kinematic solver) and the equally conditioned
mechanism (naive dynamics), indicating suboptimal solver ∆t = 1, Kp = diag([1, 1, 1, 0.1, 0.1, 0.1]),
(18)
convergence if they were applied to solve IK. Note that us- Kd = 0, K = αI6 ,
ing dynamics does not necessarily improve homogenization
where α is a scaling factor built from the mean matrices
(middle column). Only with the end effector approach (right
from Fig. 4b (top row, left + right)
column) do we effectively obtain the intended behavior.
We propose the values from (17) to be considered as good mean(diagonal(J H −1 J T ))
α= ≈ 0.7885 (19)
default values for a broad usage. We also take these values mean(diagonal(J J T ))
for the following experiments of this paper.
to obtain an identical mapping f → ẍ (considering the
B. Solver Convergence average of the main diagonal). Fig. 5 shows the systems’ step
In this experiment we analyze our control scheme from response to a sudden Cartesian offset. Both systems approach
Fig. 1 for a distant IK target, using both the Jacobian trans- the goal state, as indicated by the vanishing errors. However,
pose method and our conditioned inertia method. The distant there is a big difference in the intermediate solutions: The
Jacobian transpose method overshoots in four out of six
2 https://github.com/ros-industrial/universal_robot dimensions, and looses track of ry , which does not flat
 0.2 0.2 0.2

εx [m]

εy [m]

εz [m]
0.0 0.0 0.0
jacobian transpose
our method
−0.2 −0.2 −0.2
0 50 100 150 0 50 100 150 0 50 100 150
iterations N

0.0 0.0 0.0

εrz [rad]
εrx [rad]

εry [rad]
−0.1 −0.1 −0.1

−0.2 −0.2 −0.2

0 50 100 150 0 50 100 150 0 50 100 150
iterations N

Fig. 5: Experiment B. Analysis of the solver convergence for the Jacobian method: q̈ = αI6 J T Kp (xd − g(q)), and our method:
q̈ = H −1 J T Kp (xd − g(q)), with a simulation time interval of ∆t = 1 s and N = 150 iterations. The plots show the six Cartesian
dimensions of the error xd − g(q) for each iteration.

experiment D with a constant speed of 0.2 m/s. During the

execution we sample the moving target with a frequency of
100 Hz. Fig. 8 shows the tracking performance for individual
gains kp , which we use as a multiplicative factor for Kp
from the previous experiments. The results show the intended
behavior: By increasing the gains we transition from a
smooth but delayed controller to a fast and exact IK solver.
As an example, users may prefer kp = 5 over kp = 50 for
smoothed corners for their robot control.
Fig. 6: Robot starting configurations and target motion for three
experiments. VI. DISCUSSION
A. Virtual Dynamics
out for the interval observed. In comparison, our proposed To obtain (9), we neglected non-linear Coriolis and gravity
method converges stronger and maintains the rotational er- terms in favor of a simple, unbiased mapping from Cartesian
rors constant throughout the path to the target. end effector offset to joint accelerations. There are, however,
scenarios in which these terms could provide beneficial
C. Interpolation Quality features. For redundant manipulators, gravity and Coriolis
In the next experiment we analyze the interpolation quality terms could be used to gain more control over links and
of intermediate states. For this purpose we interpolated to joints in the nullspace, e.g. to implement a natural behavior
four targets, representing the corners of a square. The starting of moving the elbow. For manipulators with six or less
configuration is illustrated in Fig. 6 for experiment C. We degrees of freedom, however, these terms add computational
chose ∆t = 0.1 s and the number of iterations N = 50 for complexity without real advantages.
each step. Starting at all four corners we computed 1000 Our algorithm does not accumulate velocity during time
steps of intermediate states, for each taking the next counter integration, such that the system always starts anew in each
clockwise corner as fixed target xd . Fig. 7 shows the results. control cycle with instantaneous accelerations, having the
The plots demonstrate that our proposed method generates benefit that no damping is required to avoid overshooting.
goal-directed, intermediate solutions. In comparison, the Ja- Also note, that the dynamic parameters from (17), which we
cobian transpose method leads to distorted paths in Cartesian recommend as default values for any robotic manipulator
space. Note also that our method converges in less iterations, using our IK algorithm, are not necessarily realistic, and
as indicated with bigger spaces between individual points. are, in fact, deliberately off from the real system’s mass and
inertia distribution. In this regard, our used term twin refers
D. Motion tracking to the robot kinematics only. The dynamics are virtually
In this experiment, we apply our method to the tracking conditioned to obtain an IK solver with the features as
of a moving target that follows the square from Fig. 6 for described in this work.
 0.75 0.70

0.70 0.65

0.65 kp = 0.5
0.60 kp = 1

z [m]
kp = 2
jacobian transpose
z [m]

0.60
0.55 kp = 5
our method
kp = 50
0.55
0.50

0.50
0.45

0.45
0.15 0.20 0.25 0.30 0.35 0.40 0.45
0.40 y [m]
0.2 0.3 0.4
y [m]

error [m]
0.425 0.1
x [m]

0.400
0.375
0.0
0 1 2 3 4 5 6
0 500 1000 1500 2000 2500 3000 3500 4000
time [s]
steps

Fig. 7: Experiment C. Analysis of path accuracy during interpo-

lation to the corners of a square. We use the parameters from (18)
error [rad]

with ∆t = 0.1 s. The Cartesian end effector positions are computed 0.001
with the robots forward kinematics g(q) for each step.
0.000
0 1 2 3 4 5 6
B. Controller Gains time [s]

We chose the controller gains empirically for our exper-

iments. The configuration from Fig. 8 represents a good Fig. 8: Experiment D. Analysis of tracking performance for
starting point for most use cases. The magnitude of the a square-moving target of 0.2 m/s. The family of curves
controller gains directly corresponds to how fine grained the is computed by varying kp in the gain matrix Kp =
solver interpolates between the sampled targets. Users can kp diag([1, 1, 1, 0.1, 0.1, 0.1]). We chose ∆t = 0.1 s and N = 10
as IK solver parameters. Plot (a) shows the end effector’s absolute
easily adapt this responsiveness according to their use case motion in Cartesian space. Plot (b) shows the according translational
through a single factor (kp from Fig. 8). Additionally, the error k[x y z ]T k and rotational error k[rx ry rz ]T k to the
relative difference between the individual elements in the moving target.
diagonal gain matrix Kp offers the possibility to adjust how
translation is weighed against orientation, and also which
Cartesian dimension gets priority in underactuated system of a virtually conditioned twin of the manipulator into the
configurations. control loop for solving IK. Deriving our concept from the
view of manipulator forward dynamics simulations, we offer
C. Efficiency an intuitive solution to achieve task space homogenization
Computing the inverse of the manipulator’s mass matrix despite varying robot joint configurations, for which we pro-
in each controller cycle introduces additional computational vided an empirical proof on the UR10 robot. We tested and
cost. However, we did not experience this to become a compared our method in experiments with the UR10 to the
performance issue, even for solver iteration rates of 10 kHz, well-known Jacobian transpose method. The results showed
which we tested on an Intel® Core™ i7-4900MQ. Experi- that our solution outperforms the Jacobian transpose method
ment C showed that our solver converges in substantially less in terms of convergence, path accuracy and interpolation
iterations. Weighing this advantage against the more complex quality. Users may configure the algorithm to anything from
computation scheme in terms of overall convergence time a smooth controller to a fast and accurate IK solver, also for
could be subject to further study. applications outside the context of robot motion tracking.

VII. CONCLUSIONS ACKNOWLEDGMENT

In this work we presented a new method for sample- This work was supported in part by the European
based motion tracking for robotic manipulators. The core Community Seventh Framework Program under Grant
concept of our approach bases on including the mass matrix no. 608849 (EuRoC Project).
 R EFERENCES acceleration control of manipulators”. In: Journal of
[1] Andreas Aristidou et al. “Inverse Kinematics Tech- Dynamic Systems, Measurement, and Control 110.1
niques in Computer Graphics: A Survey”. In: Com- (1988), pp. 31–38.
puter Graphics Forum. Vol. 37. 6. Wiley Online [8] Patrick Beeson and Barrett Ames. “TRAC-IK: An
Library. 2018, pp. 35–58. Open-Source Library for Improved Solving of Generic
[2] Rosen Diankov et al. “Manipulation planning for the Inverse Kinematics”. In: IEEE-RAS International Con-
jsk kitchen assistant robot using openrave”. In: The ference on Humanoid Robots (Humanoids). Seoul,
29th Annual Conference on Robotics Society of Japan, Korea, Nov. 2015.
AC2Q2–2. 2011. [9] Samuel R Buss and Jin-Su Kim. “Selectively damped
[3] William A Wolovich and H Elliott. “A computational least squares for inverse kinematics”. In: Journal of
technique for inverse kinematics”. In: The 23rd IEEE Graphics tools 10.3 (2005), pp. 37–49.
Conference on Decision and Control. 1984, pp. 1359– [10] Morgan Quigley et al. “ROS: an open-source Robot
1363. Operating System”. In: ICRA workshop on open
[4] A Balestrino, G De Maria, and L Sciavicco. “Robust source software. Vol. 3. 3.2. 2009, p. 5.
control of robotic manipulators”. In: IFAC Proceed- [11] Sachin Chitta et al. “ros_control: A generic and simple
ings Volumes 17.2 (1984), pp. 2435–2440. control framework for ROS”. In: The Journal of Open
[5] Alexandre N Pechev. “Inverse kinematics without Source Software (2017).
matrix inversion”. In: IEEE International Conference [12] John J Craig. Introduction to robotics: mechanics and
on Robotics and Automation (ICRA). 2008, pp. 2005– control, 3/E. Pearson Education India, 2009.
2012. [13] S. Scherzinger, A. Roennau, and R. Dillmann. “For-
[6] Yoshihiko Nakamura and Hideo Hanafusa. “Inverse ward Dynamics Compliance Control (FDCC): A new
kinematic solutions with singularity robustness for approach to cartesian compliance for robotic manipu-
robot manipulator control”. In: ASME, Transactions, lators”. In: IEEE/RSJ International Conference on In-
Journal of Dynamic Systems, Measurement, and Con- telligent Robots and Systems (IROS). 2017, pp. 4568–
trol 108 (1986), pp. 163–171. 4575.
[7] CW Wampler and LJ Leifer. “Applications of damped
least-squares methods to resolved-rate and resolved-