Ch.

7: Independent Joint Control

1. SISO overview

• Typical single input, single output (SISO) system:

• We want the robot tracks the desired trajectory and rejects external
disturbances
• We already have the desired trajectory, and we assume that we can
measure the actual trajectories
• Thus the first thing we need is a system description

2
 1. SISO overview

• Need a convenient input-output description of a SISO system

• Two typical representations for the plant:
– Transfer function
– State-space
• Transfer functions represent the system dynamics in terms of the
Laplace transform of the ODEs that represent the system dynamics
• For example, if we have a 1DOF system described by:
τ (t ) = Jθ(t ) + Bθ(t )
• We want the representation in the Laplace domain:
τ (s ) = s 2Jθ (s ) + sBθ (s )
= s (sJ + B )θ (s )
• Therefore, we give the transfer function as:
θ (s )
P (s ) ≡
1 1/ J
= =
τ (s ) s (sJ + B ) s (s + B / J ) 3

1. SISO overview

• A generic 2nd order system can be described by the following ODE:

mx(t ) + bx (t ) + kx (t ) = F (t )
• And we want to get a transfer function representation of the system,
take the Laplace transform of both sides:
mL{x(t )} + bL{x (t )} + kL{x (t )} = L{F (t )}
( )
m s 2 x (s ) − sx (0 ) − x (0 ) + b(sx (s ) − x (0 )) + kx (s ) = F (s )
( )
ms 2 + bs + k x (s ) = F (s ) + mx (0 ) + (ms + c )x (0 )
• Ignoring the transient response, we can rearrange:
x (s ) 1
=
F (s ) ms + bs + k
2

• This is the input-output transfer function and the denominator is

called the characteristic equation

4
 2. Motor dynamics

• DC motors are ubiquitous in robotics applications

• We develop a transfer function that describes the relationship
between the input voltage and the output angular displacement
• Physical description of the most common motor: permanent
magnet…

stator

rotor
(armature)

commutator
5

2. Motor dynamics

6
 2. Motor dynamics

• When a conductor moves in a magnetic field, a voltage is generated

– The magnitude of torque:
τ m = K1φ ia
– Called back EMF:
Vb = K 2φωm ωm : rotor angular velocity
– Differential equation for the armature current
di a
L + Ri a = V − Vb R: armature resistance
dt L: armature inductance

2. Motor dynamics

• Since this is a permanent

magnet motor, the magnetic
flux is constant, we can write:
τ m = K1φi a = K m i a τ0 is the blocked torque
Km: torque constant when motor is stalled

• Similarly:
dθ
Vb = K 2φωm = K b m
dt
Kb: back EMF constant

• Km and Kb are numerically

equivalent, thus there is one
constant needed to
Torque-speed curves
characterize a motor

8
 2. Motor dynamics

• Let’s take motor and connect it to a link

• Between the motor and link there is a gear such that: θ m = rθ L
• Lump the actuator and gear inertias: J m = J a + J g
• Now we can write the dynamics of this mechanical system:
d 2θ m dθ τ τ
Jm 2
+ Bm m = τ m − L = K m i a − L
dt dt r r

2. Motor dynamics

• Now we have the ODEs describing this system in both the electrical
and mechanical domains:
di a dθ d 2θ m dθ τ
L + Ri a = V − K b m Jm + Bm m = K m i a − L
dt dt dt 2
dt r
• In the Laplace domain:
τ L (s )
(Ls + R )Ia (s ) = V (s ) − K bsΘ m (s ) (J m )
s 2 + Bm s Θ m (s ) = K mIa (s ) −
r
• These two can be combined to define the input-output relationship
for the input voltage, load torque, and output displacement:

10
 2. Motor dynamics

• Remember, we want to express the system as a transfer function

from the input to the output angular displacement
– But we have two potential inputs: the load torque and the
armature voltage
– First, assume τL = 0 and solve for Θm(s):

(J m )
s 2 + Bm s Θ m (s )
= Ia (s )
Km

(Ls + R )(J ms 2 + Bms )Θ (s ) = V (s ) − K sΘ (s )

m b m
Km
Θ m (s ) Km
=
V (s ) s[(Ls + R )(J m s + Bm ) + K b K m ]

11

2. Motor dynamics

• Now consider that V(s) = 0 and solve for Θm(s):

− K b sΘ m (s )
Ia (s ) =
Ls + R
− K m K b sΘ m (s ) τ L (s )
(J m )
s 2 + Bm s Θ m (s ) =
Ls + R
−
r
Θ m (s ) − (Ls + R ) / r
=
τ L (s ) s[(Ls + R )(J m s + Bm ) + K bK m ]

• Note that this is a function of the gear ratio

– The larger the gear ratio, the less effect external torques have
on the angular displacement

12
 2. Motor dynamics

• In this system there are two ‘time constants’

– Electrical: L/R
– Mechanical: Jm/Bm
• For intuitively obvious reasons, the electrical time constant is
assumed to be small compared to the mechanical time constant
– Thus, ignoring electrical time constant will lead to a simpler
version of the previous equations:

Θ m (s ) Km / R
=
V (s ) s[J m s + Bm + K bK m / R ]

Θ m (s ) − 1/ r
=
τ L (s ) s[J m s + Bm + K bK m / R ]

13

2. Motor dynamics

• Rewriting these in the time domain gives:

Θ m (s ) Km / R J mθm (t ) + (Bm + K b K m / R )θm (t ) = (K m / R )V (t )
=
V (s ) s[J m s + Bm + K bK m / R ]
Θ m (s ) − 1/ r J mθm (t ) + (Bm + K b K m / R )θm (t ) = −(1/ R )τ L (t )
=
τ L (s ) s[J m s + Bm + K bK m / R ]

• By superposition of the solutions of these two linear 2nd order ODEs:

J mθm (t ) + (Bm + K b K m / R )θm (t ) = (K m / R )V (t ) − (1/ R )τ L (t )

J B u (t ) d (t )
• We can write the dynamics of a DC motor attached to a load as:
Jθ( t ) + Bθ ( t ) = u ( t ) − d ( t ) ( )
Js 2 + Bs Θ ( s ) = U ( s ) − D ( s )
– Note that u(t) is the input and d(t) is the disturbance (e.g. the
dynamic coupling from motion of other links)
14
 3. Set-point Tracking

• We will first discuss two initial controllers: PD and PID

– Both attempt to drive the error (between a desired trajectory
and the actual trajectory) to zero
• The system can have any dynamics, but we will concentrate on
the previously derived system

15

3.1. PD controller

• Control law: u (t ) = K pe(t ) + K d e (t )

– Where e(t) = θd(t) - θ(t)
• In the Laplace domain: U (s ) = (K p + sK d )E (t )
(
U (s ) = (K p + sK d ) Θ d (s ) − Θ (s ))
Θ(s) =
U (s) − D(s)
=
( K p + sK d ) Θd ( s ) − D ( s )
Js 2 + Bs Js 2 + ( B + K d ) s + K p

• The closed-loop system:

16
 3.1. PD controller

• The denominator is the characteristic polynomial

• The roots of the characteristic polynomial determine the
performance of the system

s2 +
(B + K d ) s + K p =0
J J
• If we think of the closed-loop system as a damped second order
system, this allows us to choose values of Kp and Kd
s 2 + 2ζωs + ω 2 = 0
• Thus Kp and Kd are:
K p = ω 2J
K d = 2ςωJ − B

• A natural choice is ζ = 1 (critically damped)

17

3.1. PD controller

• Limitations of the PD controller:

– for illustration, let our desired trajectory be a step input and our
disturbance be a constant as well:
Θ d (s ) = , D(s ) =
C D
s s
– Plugging this into our system description gives:
Θ (s ) =
(K p + sK d )C − D
(
s Js + (B + K d )s + K p
2
)
– The steady-state value of the displacement:
s (K p + sK d )C − sD (K + sK d )C − D K pC − D D
θ ss = lim = lim p
= =C −
s →0 (
s Js + (B + K d )s + K p
2
) s →0 Js + (B + K d )s + K p
2
Kp Kp

– Thus the steady state error is: –D/Kp

– Therefore to drive the error to zero in the presence of large
disturbances, we need large gains.
18
 3.2. PID controller

• Control law: u (t ) = K pe(t ) + K d e (t ) + K i ∫ e(t )dt

• In the Laplace domain:

⎛ K ⎞
U (s ) = ⎜ K p + K d s + i ⎟E (s )
⎝ s ⎠

Θ (s ) =
(K sd
2
)
+ K ps + K i Θ d (s ) − sD(s )
Js + (B + K d )s 2 + K ps + K i
3

19

3.2. PID controller

• The integral term eliminates the steady state error that can arise from
a large disturbance
• How to determine PID gains
1. Set Ki = 0 and solve for Kp and Kd
(B + K d )K p
2. Determine Ki to eliminate steady state error K i <
J
• Stability
– The closed-loop stability of these systems is determined by the
roots of the characteristic polynomial
– If all roots (potentially complex) are in the ‘left-half’ plane, our
system is stable
– A description of how the roots of the characteristic equation
change (as a function of controller gains) is very valuable
• Called the root locus
20
 4. Trajectory tracking

• Setpoint controllers are inappropriate for tracking time-varying

trajectories
• Thus we need alternative controllers
– Feedforward control
– Computed torque control
– State feedback control

21

4.1. Feedforward control

• If we want to track an arbitrary trajectory, and we know the system

dynamics, we can use the following architecture:

– Assumptions:
• G(s), H(s) are proper
• G(s) is open-loop stable
• G(s), H(s), F(s) are rational

22
 4.1. Feedforward control

• Since the three polynomials are rational, we can write them as:
q (s ) c (s ) a(s )
G(s ) = , H (s ) = , F (s ) =
p(s ) d (s ) b(s )
• Where:
deg q (s ) < deg p(s ), degc (s ) < deg d (s )

• Again, call u(t) the input to the open loop plant:

U (s ) = F (s )Θ d + H (s )(Θ d − Θ )

• Next we can write the output:

Θ = G(s )[F (s )Θ d + H (s )(Θ d − Θ )]
• Solving this for Θ gives:

Θ=
(G(s )F (s ) + G(s )H (s ))Θ d
1 − G(s )H (s )
23

4.1. Feedforward control

• Rewriting this in terms of the rational functions gives:

T (s) =
Θ(s)
=
(G ( s ) F ( s ) + G ( s ) H ( s ))
Θd ( s ) 1− G (s) H (s)
q ( s ) ⎡⎣c ( s ) b ( s ) + a ( s ) d ( s ) ⎤⎦
=
b ( s ) ⎡⎣ p ( s ) d ( s ) + q ( s ) c ( s ) ⎤⎦

• Thus, if the feedforward system is to be stable, the closed loop

characteristic polynomial has to have poles in the left-half plane
– Closed-loop characteristic polynomial
– Therefore: b(s )[p(s )d (s ) + q (s )c (s )]
Re(roots(b(s ))) < 0
Re(roots(p(s )d (s ) + q (s )c (s ))) < 0

24
 4.1. Feedforward control

• What is a good candidate for F(s)?

• Since we know the open loop dynamics G(s), choose F(s) = 1/G(s)
– i.e. a(s) = p(s), b(s) = q(s)
• Now we can rewrite the closed-loop transfer function:
Θ ( s ) q ( s ) ⎡⎣c ( s ) b ( s ) + a ( s ) d ( s ) ⎤⎦
T (s) = d =
Θ ( s ) b ( s ) ⎣⎡ p ( s ) d ( s ) + q ( s ) c ( s ) ⎦⎤
q ( s ) ⎡⎣c ( s ) q ( s ) + p ( s ) d ( s ) ⎤⎦
= =1
q ( s ) ⎡⎣ p ( s ) d ( s ) + q ( s ) c ( s ) ⎤⎦
• i.e. the tracking error is zero
• But we need to be careful: we can only do this when the plant is
minimum phase
– i.e. all zeros are in the left-half plane

25

4.1. Feedforward control

• If there is a disturbance, our system looks like:

• The transfer function:

(
Θ ( s ) = G ( s ) ⎡ D ( s ) + F ( s ) Θd ( s ) + H ( s ) Θd ( s ) − Θ ( s ) ⎤
⎣ ⎦ )
Θ ( s ) = G ( s ) ⎡ D ( s ) + F ( s ) Θd
⎣
( s) + H ( s)(Θ d
( s ) − Θ ( s ) )⎤⎦
= G ( s ) D ( s ) + G ( s ) F ( s ) Θd ( s ) + G ( s ) H ( s ) Θd ( s ) − G ( s ) H ( s ) Θ ( s )
G (s) G (s) F (s) + G (s) H (s) d
= D(s) + Θ (s)
1+ G (s) H (s) 1+ G (s) H (s)
26
 4.1. Feedforward control

• Again defining the error as E = Θd - Θ gives us

E ( s ) = Θd ( s ) − Θ ( s )
G (s) G (s) F (s) + G (s) H (s) d
= Θd ( s ) − D(s) − Θ (s)
1+ G (s) H (s) 1+ G (s) H (s)
G (s) 1− G (s) F (s) d
=− D(s) + Θ (s)
1+ G (s) H (s) 1+ G (s) H (s)

• Since we defined F(s) = 1/G(s), the last term is zero

G (s)
E (s) = − D(s)
1+ G (s) H (s)

27

4.1. Feedforward control

• Example: Consider the system with the previous dynamics and a PD

compensator

G(s ) = F (s ) = Js 2 + Bs
1
H (s ) = K p + K d s
Js + Bs
2

28
 4.1. Feedforward control

• Computed torque control

29

4.2. State space representations

• Another representation of dynamical systems allows us to use a

matrix form
• Consider the (open-loop) system described by the following ODE:
τ = Jθ + Bθ + Kθ
• Now choose:
x1 = θ
x 2 = x1 = θ
• We call x1 and x2 the states of our system, and we can now rewrite
our ODE in matrix form:
x = Ax + bu
⎡ 0 1 ⎤ ⎡0⎤
⎡ x ⎤
= ⎢ k −B ⎥ ⎢ ⎥ + ⎢ 1 ⎥ u
1
⎢− ⎥ ⎣ x2 ⎦ ⎢ ⎥
⎣ J J ⎦ ⎣J ⎦
30
 4.2. State space representations

• This gives a two dimensional set of first order ODEs describing the
evolution of the state
• If we are interested in an output (for example θ), we make one last
definition:
y = θ = x1 = c T x
= [1 0]x

• What is the relationship to the transfer function representation?

– Calling u(t) the input, y(t) the output,
Y (s )
G(s ) = = c T (sI − A ) b
−1

U (s )

31

4.2. State space representations

• Further, the roots of the characteristic polynomial are the

eigenvalues of A
• State feedback control
– Control law: n
u (t ) = −k T x = ∑ k i xi
i =1

– Where the ki are gains to be determined

– i.e. the control is a linear combination of the states (typically
positions and velocities)
– Substituting this into our state space representation gives
( )
x = Ax + bu = Ax − bk T x = A − bk T x
– Thus, state feedback changes the eigenvalues of the closed-
loop system (i.e. the poles of the closed loop system)

32