Robotics 2

Linear parametrization and

identification of robot dynamics

Prof. Alessandro De Luca

 Dynamic parameters of robot links

n consider a generic link Center of Mass

of a fully rigid robot (CoM) frame 𝑖
link 𝑖 𝐼+' 𝑧𝑖

joint 𝑖
𝑟+' 𝑦𝑖 kinematic frame 𝑖
CoM 𝐶𝑖 (DH or modified DH)
𝑚' 𝑔) 𝑂𝑖
𝑥𝑖
base 𝑟'
frame 0
𝑟/' 𝐼+',// 𝐼+',/0 𝐼+',/1
n each link is characterized by 𝑚' 𝑟+' = 𝑟0' 𝐼+',00 𝐼+',01
𝐼+' =
10 dynamic parameters 𝑟1' symm 𝐼+',11
n however, the robot dynamics depends in a nonlinear way on some of these
8
parameters (e.g., through the combination 𝐼6',11 + 𝑚' 𝑟/' )
Robotics 2 2
 Dynamic parameters of robots
n kinetic energy and gravity potential energy can both be rewritten so
that a new set of dynamic parameters appears only in a linear way
§ need to re-express link inertia and CoM position in (any) known kinematic
frame attached to the link (same orientation as the barycentric frame)
n fundamental kinematic relation
𝑣6' = 𝑣' + 𝜔' × 𝑟+' = 𝑣' + 𝑆 𝜔' 𝑟+' = 𝑣' − 𝑆 𝑟+' 𝜔'
n kinetic energy of link 𝑖
1 A 1 A
𝑇' = 𝑚' 𝑣+' 𝑣+' + 𝜔' 𝐼+' 𝜔'
2 2
B B
= 𝑚' 𝑣' − 𝑆 𝑟+' 𝜔' A 𝑣' − 𝑆 𝑟+' 𝜔' + 𝜔'A 𝐼+' 𝜔'
8 8
1 1 A
= 𝑚' 𝑣' 𝑣' + 𝜔' 𝐼+' + 𝑚' 𝑆 A 𝑟+' 𝑆 𝑟+' 𝜔' − 𝑣'A 𝑆 𝑚' 𝑟+' 𝜔'
A
2 2
𝐼',// 𝐼',/0 𝐼',/1
Steiner theorem 𝐼' = 𝐼',00 𝐼',01
symm 𝐼',11
Robotics 2 3
 Standard dynamic parameters of robots
n gravitational potential energy of link 𝑖
𝑈' = −𝑚' 𝑔)A 𝑟),+' = −𝑚' 𝑔)A 𝑟' + 𝑟+' = −𝑚' 𝑔)A 𝑟' − 𝑔)A 𝑚' 𝑟+'
n by expressing vectors and matrices in frame 𝑖, both 𝑇' and 𝑈' are
linear in the set of 10 (constant) standard parameters 𝜋' ∈ ℝB)
1 A𝑖 A 1𝑖 A 𝑖 𝑖
𝑇' = 𝑚' 𝑣' 𝑣' + 𝑚' 𝑟+' 𝑆 𝑣' 𝜔' + 𝜔' 𝐼' 𝜔'
𝑖 𝑖 𝑖 𝑖
2 2
𝑈' = −𝑚' 𝑔)A 𝑟' − 𝑔)A 0𝑅' (𝑚' 𝑖 𝑟+' )
mass×CoM
mass of link 𝑖 position of link 𝑖 inertia of link 𝑖
(0-th order (1-st order (2-nd order
moment ) moment )
moment )
𝑚'
𝝅𝒊 = 𝑚' 𝑖𝑟+' = 𝑚' 𝑚' 𝑖𝑟+',/ 𝑚' 𝑖𝑟+',0 𝑚' 𝑖𝑟+',1 𝑖𝐼
',//
𝑖𝐼
',/0
𝑖𝐼
',/1
𝑖𝐼
',00
𝑖𝐼
',01
𝑖𝐼
',11
A
𝑣𝑒𝑐𝑡 𝑖 𝐼'

n since the E-L equations involve only linear operations on 𝑇 and 𝑈, also
the robot dynamic model is linear in the standard parameters 𝝅 ∈ ℝB)G
Robotics 2 4
 Linearity in the dynamic parameters
n using a 𝑁 × 10𝑁 regression matrix 𝑌S that depends only on kinematic
quantities, the robot dynamic equations can always be rewritten linearly
in the standard dynamic parameters as
𝑀 𝑞 𝑞̈ + 𝑐 𝑞, 𝑞̇ + 𝑔 𝑞 = 𝑌S 𝑞, 𝑞,̇ 𝑞̈ 𝜋 = 𝑢
𝜋 A = 𝜋BA 𝜋8A ⋯ A
𝜋G
n the open kinematic chain structure of the manipulator implies that the 𝑖 -th
dynamic equation can depend only on the standard dynamic parameters
of links 𝑖 to 𝑁 ⇒ 𝑌S has a block upper triangular structure
𝑌BB 𝑌B8 ⋯ 𝑌BG
0 𝑌88 ⋯ 𝑌8G with row vectors
𝑌S 𝑞, 𝑞,̇ 𝑞̈ = 𝑌',[ of size 1×10
⋮ ⋱ ⋮
0 ⋯ 0 𝑌GG
Property: element 𝑚'[ of 𝑀(𝑞) is a function at most of (𝑞]^B , ⋯ , 𝑞G ), for 𝑘 = min{𝑖, 𝑗} ,
and of the inertial parameters of at most links 𝑟 to 𝑁 , with 𝑟 = max{𝑖, 𝑗}
Robotics 2 5
 Linearity in the dynamic coefficients
n many standard parameters do not appear (“play no role”) in the dynamic model of a
given robot ⇒ the associated columns of 𝑌S are 0!
n some standard parameters may appear only in fixed combinations with others ⇒ the
associated columns of 𝑌S are linearly dependent!
n one can isolate 𝑝 ≪ 10𝑁 independent groups of parameters 𝜋 (associated to 𝑝
functionally independent columns 𝑌'jklm of 𝑌S ) and partition matrix 𝑌S in two
blocks, the second containing dependent (or zero) columns as 𝑌klm = 𝑌'jklm 𝑇, for a
suitable constant 𝑝 × (10𝑁 − 𝑝) matrix 𝑇
𝜋'jklm 𝜋'jklm
𝑌S 𝑞, 𝑞,̇ 𝑞̈ 𝜋 = 𝑌'jklm 𝑌klm 𝜋klm = 𝑌'jklm 𝑌'jklm 𝑇 𝜋klm
= 𝑌'jklm 𝜋'jklm + 𝑇 𝜋klm = 𝑌 𝑞, 𝑞,̇ 𝑞̈ 𝑎

n these grouped parameters are called dynamic coefficients 𝑎 ∈ ℝm , “the only that
matter” in robot dynamics (= base parameters by W. Khalil)
n the minimal number 𝑝 of dynamic coefficients that is needed can also be checked
numerically (see later → Identification)
Robotics 2 6
 Linear parameterization of robot dynamics
it is always possible to rewrite the dynamic model in the form
regression 𝑎 = vector of
matrix dynamic coefficients

𝑀 𝑞 𝑞̈ + 𝑐 𝑞, 𝑞̇ + 𝑔 𝑞 = 𝑌 𝑞, 𝑞,̇ 𝑞̈ 𝑎 = 𝑢
𝑁×𝑝 𝑝×1
e.g., the heuristic grouping (found by inspection) on a 2R planar robot
𝑎B = 𝐼6B,11 + 𝑚B 𝑑B8 + 𝐼68,11 + 𝑚8 𝑑88 + 𝑚8 𝑙B8
𝑎B 𝑎8 = 𝑚8 𝑙B 𝑑8
𝑎8 𝑎s = 𝐼68,11 + 𝑚8 𝑑88
𝑞̈ B 𝑐8 2𝑞̈ B + 𝑞̈ 8 − 𝑠8 𝑞̇ 88 + 2𝑞̇ B 𝑞̇ 8 𝑞̈ 8 𝑐B 𝑐B8
𝑎s =
𝑢B
0 𝑐8 𝑞̈ B + 𝑠8 𝑞̇ B8 𝑞̈ B + 𝑞̈ 8 0 𝑐B8 𝑢8 𝑎u = 𝑔) 𝑚B 𝑑B + 𝑚8 𝑙B
𝑎u
𝑎t 𝑎t = 𝑔) 𝑚8 𝑑8
NOTE: 4 more coefficients are added when including the coefficients 𝐹p,' and 𝐹+,' of viscous and
Coulomb friction at the joints (“decoupled” terms appearing only in the relative equations 𝑖 = 1,2)
Robotics 2 7
 Linear parametrization
of a 2R planar robot (𝑁 = 2)

n being the kinematics known (i.e., 𝑙B and 𝑔) ), the number of dynamic

coefficients can be reduced since we can merge the two coefficients
𝑎8 = 𝑚8 𝑙B 𝑑8 & 𝑎t = 𝑔) 𝑚8 𝑑8 ⇒ 𝑎8 = 𝑚8 𝑑8 (factoring out 𝑙B and 𝑔) )
n therefore, after regrouping, 𝒑 = 𝟒 dynamic coefficients are sufficient
𝑎B
𝑞̈ B 𝑙B 𝑐8 2𝑞̈ B + 𝑞̈ 8 − 𝑙B 𝑠8 𝑞̇ 88 + 2𝑞̇ B 𝑞̇ 8 + 𝑔) 𝑐B8 𝑞̈ 8 𝑔) 𝑐B 𝑎8 𝑢B
𝑎s = 𝑌 𝑎 = 𝑢 = 𝑢8
0 𝑙B 𝑐8 𝑞̈ B + 𝑠8 𝑞̇ B8 + 𝑔) 𝑐B8 𝑞̈ B + 𝑞̈ 8 0
𝑎u
𝑎B = 𝐼6B,11 + 𝑚B 𝑑B8 + 𝐼68,11 + 𝑚8 𝑑88 + 𝑚8 𝑙B8 𝑎s = 𝐼68,11 + 𝑚8 𝑑88
𝑎8 = 𝑚8 𝑑8 𝑎u = 𝑚B 𝑑B + 𝑚8 𝑙B

n this (minimal) linear parametrization of robot dynamics is not unique,

both in terms of the chosen set of dynamic coefficients 𝑎 and for the
associated regression matrix 𝑌
n a systematic procedure for its derivation would be preferable

Robotics 2 8
 Linear parametrization
of a 2R planar robot (𝑁 = 2)
n as alternative to the previous heuristic method, apply the general procedure
n 10𝑁 = 20 standard parameters are defined for the two links

n from the assumptions made on CoM locations, only 5 such parameters actually

appear, namely (with 𝑑' = 𝑟6',/ )

link 1: 𝑚B 𝑑B 𝐼B,11 = 𝐼6B,11 + 𝑚B 𝑑B8 link 2: 𝑚8 𝑚8 𝑑8 𝐼8,11 = 𝐼68,11 + 𝑚8 𝑑88
n in the 2×5 matrix 𝑌S , the 3rd column (associated to 𝑚8 ) is 𝑌Ss = 𝑌SB 𝑙B + 𝑌S8 𝑙B8
n after regrouping/reordering, 𝒑 = 𝟒 dynamic coefficients are again sufficient
𝑎B
8 𝑢B
𝑔) 𝑐B 𝑞̈ B 𝑙B 𝑐8 2𝑞̈ B + 𝑞̈ 8 − 𝑙B 𝑠8 𝑞̇ 8 + 2𝑞̇ B 𝑞̇ 8 + 𝑔) 𝑐B8 𝑞̈ B + 𝑞̈ 8 𝑎8
= 𝑌 𝑎 = 𝑢 =
0 0 𝑙B 𝑐8 𝑞̈ B + 𝑠8 𝑞̇ B8 + 𝑔) 𝑐B8 𝑞̈ B + 𝑞̈ 8 𝑎s 𝑢8
𝑎u
8 8 8 𝑎s = 𝑚8 𝑑8
𝑎B = 𝑚B 𝑑B + 𝑚8 𝑙B 𝑎8 = 𝐼B,11 + 𝑚8 𝑙B = 𝐼6B,11 + 𝑚B 𝑑B + 𝑚B 𝑙B
𝑎u = 𝐼8,11 = 𝐼68,11 + 𝑚8 𝑑88
n determining a minimal parameterization (i.e., minimizing 𝑝) is important for
n experimental identification of dynamic coefficients
n adaptive/robust control design in the presence of uncertain parameters
Robotics 2 9
 Identification of dynamic coefficients
n in order to “use” the model, one needs to know the numeric values of
the robot dynamic coefficients
n robot manufacturers provide at most only a few principal dynamic

parameters (e.g., link masses)

n estimates can be found with CAD tools (e.g., assuming uniform mass)
n friction coefficients are (slowly) varying over time
n lubrication of joints/transmissions

n for an added payload (attached to the E-E)

n a change in the 10 dynamic parameters of last link

n this implies a variation of (almost) all robot dynamic coefficients

n preliminary identification experiments are needed

n robot in motion (dynamic issues, not just static or geometric ones!)

n only the robot dynamic coefficients can be identified (and not all

the link standard parameters!)

Robotics 2 10
 Identification experiments
1. choose a motion trajectory 𝑞k (𝑡) that is sufficiently “exciting”, i.e.,
§ explores the robot workspace and involves all components in the
robot dynamic model
§ is periodic, with multiple frequency components
2. execute this motion (approximately) by means of a control law
§ taking advantage of any available information on the robot model
§ often 𝑢 = 𝐾{ (𝑞k − 𝑞) + 𝐾| (𝑞̇ k − 𝑞)
̇ (PD, no model information used)
3. measure 𝑞 (encoders) in 𝑛6 time instants (and, if available, also 𝑞̇ )
§ joint velocity 𝑞̇ and acceleration 𝑞̈ can be later estimated off line by
numerical differentiation (use of non-causal filters is feasible)
4. with such measures/estimates, evaluate the regression matrix 𝑌 (on the
left) and use the applied commands 𝑢 (on the right) in the expression

𝑌 𝑞 𝑡] , 𝑞̇ 𝑡] , 𝑞̈ 𝑡] 𝑎 = 𝑢 𝑡] 𝑘 = 1, ⋯ , 𝑛6
Robotics 2 11
 Least Squares (LS) identification
n set up the system of linear equations
𝑌 𝑞 𝑡B , 𝑞̇ 𝑡B , 𝑞̈ 𝑡B 𝑢 𝑡B
𝑛6 × 𝑁 ⋮ 𝑎= ⋮ • = 𝑢•
𝑌𝑎
𝑌 𝑞 𝑡j‚ , 𝑞̇ 𝑡j‚ , 𝑞̈ 𝑡j‚ 𝑢 𝑡j‚

n sufficiently “exciting” trajectories, large enough number of

samples (𝑛6 × 𝑁 ≫ 𝑝), and their suitable selection/position,
guarantee rank(𝑌 • ) = 𝑝 (full column rank)
n solution by pseudoinversion of matrix 𝑌 •
𝑎 = 𝑌• # 𝑢• = 𝑌• A 𝑌• •B 𝑌• A 𝑢• (∈ ℝm )
n one can also use a weighted pseudoinverse, to take into
account different levels of noise in the collected measures

Robotics 2 12
 Additional remarks on LS identification
n it is convenient to preserve the block (upper) triangular structure of
the regression matrix, by “stacking” all time evaluations in row by row
sequence of the original 𝑌 matrix
𝑌B 𝑞 𝑡B , 𝑞̇ 𝑡B , 𝑞̈ 𝑡B
𝑛6
𝑛6 ⋮ 𝑢B 𝑡B
⋮
𝑌B 𝑞 𝑡j‚ , 𝑞̇ 𝑡j‚ , 𝑞̈ 𝑡j‚ 𝑢B 𝑡j‚
𝑌8 𝑞 𝑡B , 𝑞̇ 𝑡B , 𝑞̈ 𝑡B 𝑢8 𝑡B 𝑌• =
𝑁× ⋮
𝑎=
⋮ • = 𝑢•
𝑌𝑎
𝑌8 𝑞 𝑡j‚ , 𝑞̇ 𝑡j‚ , 𝑞̈ 𝑡j‚ 𝑢8 𝑡j‚
⋮ ⋮
𝑌G 𝑞 𝑡B , 𝑞̇ 𝑡B , 𝑞̈ 𝑡B 𝑢G 𝑡B
𝑛6 ⋮
⋮ 𝑛6
𝑢G 𝑡j‚
𝑌G 𝑞 𝑡j‚ , 𝑞̇ 𝑡j‚ , 𝑞̈ 𝑡j‚
§ numerical check of full column rank is
n further practical hints more robust ⇔ rank = 𝑝 (# of col’s)

n outlier data can be eliminated in advance (when building 𝑌 )

n if sufficiently rich friction models are not included in 𝑌𝑎 , discard the
data collected at joint velocities close to zero
Robotics 2 13
 Summary on dynamic identification
KUKA IR 361
robot and
optimal
excitation
trajectory

J. Swevers, W. Verdonck, and J. De Schutter:

“Dynamic model identification for industrial robots”
IEEE Control Systems Mag., Oct 2007
results after identification (first three joints only)
Robotics 2 14
 Dynamic identification of KUKA LWR4
video

data acquisition for identification

dynamic coefficients: 30 inertial, 12 for gravity
validation after identification (for all 7 joints):
C. Gaz, F. Flacco, A. De Luca: on new desired trajectories, compare
“Identifying the dynamic model used by the KUKA LWR:
torques computed with the identified model
A reverse engineering approach” and torques measured by joint torque sensors
IEEE ICRA 2014
Robotics 2 15
 Identification of LWR4 gravity terms
using the linear parametrization, gravity terms can also be identified separately

symbolic expressions of gravity-

related dynamic coefficients

numerical values
identified through
experiments

gravity joint torques

prediction/evaluation on
new validation trajectory
Robotics 2 16
 Role of friction in identification
KUKA LWR4 dynamic model estimation vs. joint torque sensor measurement
- friction meas

without the use of a joint friction model including an identified joint friction model

Robotics 2 17
 Dynamic identification of KUKA LWR4
video

using more dynamic robot motions for model identification

J. Hollerbach, W. Khalil, M. Gautier: “Ch. 6: Model Identification”, Springer Handbook of Robotics (2nd Ed), 2016
free access to multimedia extension: http://handbookofrobotics.org
Robotics 2 18
 Adding a payload to the robot

n in several industrial applications, changes in the

robot payload are often needed
n using different tools for various technological operations
such as polishing, welding, grinding, ...
n pick-and-place tasks of objects having unknown mass

n what is the rule of change for dynamic parameters

when there is an additional payload?
n do we obtain again a linearly parameterized problem?
n does this property rely on some specific choice of
reference frames (e.g., conventional or modified D-H)?

Robotics 2 19
 Rule of change in dynamic parameters
n only the dynamic parameters of the link where a load is added will change
(typically, added to the last one –link n– as payload)
n last link dynamic parameters: mn (mass), cn = (cnxcnycnz)T (center of mass),
In (inertia tensor expressed w.r.t. frame n)
n payload dynamic parameters: mL (mass), cL = (cLxcLycLz)T (center of
mass), IL (inertia tensor expressed w.r.t. frame n)

n mass

n center of mass
(weighted average) where i = x, y, z
valid only if tensors are expressed w.r.t.
n inertia tensor the same reference frame (i.e., frame n)!

§ linear parametrization is preserved with any kinematic convention

(the parameters of the last link will always appear in the form shown above)

Robotics 2 20
 Example: 2R planar robot with payload
y2
x2

y1 a2

y0 q2
a1 x1 g0 = gravity acceleration
q1
x0

robot dynamics robot dynamics

Note 1: position of the center of mass of the two links and of the payload may also be asymmetric
Note 2: link inertia & center of mass are expressed in the DH kinematic frame attached to the link
(e.g., I2zz is the inertia of the second link around the axis z2)
Robotics 2 21
 Validation on the KUKA LWR4 robot
video

C. Gaz, A. De Luca: “Payload estimation based on identified coefficients of robot dynamics

– with an application to collision detection” IEEE IROS 2017, Vancouver, September 2017

see the block

Robotics 2 of slides! 22
 Bibliography
n J. Swevers, W. Verdonck, J. De Schutter, “Dynamic model identification for industrial robots,” IEEE Control
Systems Mag., vol. 27, no. 5, pp. 58–71, 2007
n J. Hollerbach, W. Khalil, M. Gautier, “Model Identification,” Springer Handbook of Robotics (2nd Ed), pp. 113-
138, 2016
n C. Gaz, F. Flacco, A. De Luca, “Identifying the dynamic model used by the KUKA LWR: A reverse engineering
approach,” IEEE Int. Conf. on Robotics and Automation, pp. 1386-1392, 2014
n C. Gaz, F. Flacco, A. De Luca, “Extracting feasible robot parameters from dynamic coefficients using nonlinear
optimization methods,” IEEE Int. Conf. on Robotics and Automation, pp. 2075-2081, 2016
n C. Gaz, A. De Luca, “Payload estimation based on identified coefficients of robot dynamics – with an
application to collision detection,” IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pp. 3033-3040, 2017
n C. Gaz, E. Magrini, A. De Luca, “A model-based residual approach for human-robot collaboration during
manual polishing operations,” Mechatronics, vol. 55, pp. 234-247, 2018
n C. Gaz, M. Cognetti, A. Oliva, P. Robuffo Giordano, A. De Luca, “Dynamic identification of the Franka Emika
Panda robot with retrieval of feasible parameters using penalty-based optimization,” IEEE Robotics and
Automation Lett., vol. 4, no. 4, pp. 4147-4154, 2019

Franka Emika
Panda (7R)
KUKA Universal Robot
LWR4 (7R) UR10 (6R)

Robotics 2 23