Basic Concepts in

Control
393R: Autonomous Robots
Peter Stone

Slides Courtesy of
Benjamin Kuipers
 Good Afternoon
Colleagues

• Are there any questions?

 Logistics
• Reading responses
• Next week’s readings - due
Monday night
– Braitenberg vehicles
– Forward/inverse kinematics
– Aibo joint modeling
• Next class: lab intro (start
here)
 Controlling a Simple
System
x˙
• Consider a simple system: = F(x,u)
– Scalar variables x and u, not vectors x
and u.
– Assume x is observable: y = G(x) = x
– Assume effect of motor command u: ∂F
>0
∂u
• The setpoint xset is the desired value.
– The controller responds to error: e = x
 xset

• The goal is to set u to reach e = 0.

 The intuition behind
control
• Use action u to push back toward
error e = 0
– error e depends on state x (via sensors
y)
• What does pushing back do?
– Depends on the structure of the system
– Velocity versus acceleration control
• How much should we push back?
– What does the magnitude of u depend on?
Car on a slope example
 Velocity or
acceleration control?
• If error reflects x, does u
affect x or x ?
• Velocity control: u  x
x˙ = (x˙) = F (x,u)
(valve fills tank) = (u)
– let x = (x)

⎛x˙ ⎞ u  x ⎛v ⎞
• Acceleration control:
(rocket) x˙ = ⎜⎜ ⎟⎟ = F (x,u) = ⎜⎜ ⎟⎟
– let x = (x v)T ⎝v˙ ⎠ ⎝u⎠
v˙ = x˙˙ = u
 The Bang-Bang
Controller
• Push back, against the direction of
the error
– with constant action u
• Error is e = x - xset
e<0 ⇒ u := on ⇒ x˙ = F (x,on) > 0
e>0 ⇒ u := off ⇒ x˙ = F (x,off ) < 0
• To prevent chatter around e = 0,
e < −ε ⇒ u := on
e > +ε ⇒ u := off
• Household thermostat. Not very
subtle.
Bang-Bang Control in
Action

– Optimal for reaching the setpoint

– Not very good for staying near it
 Hysteresis
• Does a thermostat work exactly
that way?
– Car demonstration
• Why not?

• How can you prevent such

frequent motor action?
• Aibo turning to ball example
 Proportional Control
• Push back, proportional to the
error.
u = −ke + ub
x˙ = F(x set ,ub ) = 0
– set ub so that
• For a linear system, we get
exponential convergence.
−α t
x(t) = Ce + x set
• The controller gain k determines
how quickly the system responds to
error.
 Velocity Control
• You want to drive your car at
velocity vset.
• You issue the motor command u =
posaccel
• You observe velocity vobs.

• Define a first-order controller:

u = −k (v obs −v )+ u
set b

– k is the controller gain.

Proportional Control
in Action

– Increasing gain approaches setpoint faster

– Can leads to overshoot, and even
instability
– Steady-state offset
 Steady-State Offset
• Suppose we have continuing
x˙ = F(x,u) + d
disturbances:

• The P-controller cannot

stabilize at e = 0.
– Why not?
 Steady-State Offset
• Suppose we have continuing disturbances:

x˙ = F(x,u) + d
• The P-controller cannot stabilize at e =
0.
– if ub is defined so F(xset,ub) = 0
– then F(xset,ub) + d  0, so the system
changes
• Must adapt ub to different disturbances
d.
 Adaptive Control
• Sometimes one controller isn’t
enough.
• We need controllers at different
u = −k P
time scales.
e + u b
u˙ b = −k I e where k I << kP

• This can eliminate steady-state

€ offset.
€ – Why?
 Adaptive Control
• Sometimes one controller isn’t
enough.
• We need controllers at different
time uscales.
= −k P e + ub
u˙ b = −k I e where k I << kP

• This can eliminate steady-state

€ offset.
€ – Because the slower controller adapts ub.
 Integral Control
u˙ b
• The adaptive controller = −k I e
means t
ub (t) = −kI ∫ e dt + ub
0

• Therefore
€ t
u(t) = −kP e(t) − k I ∫ e dt + u b
0
• The Proportional-Integral (PI)
Controller.
 Nonlinear P-control
• Generalize proportional control to
+
u = − f (e) + ub where f ∈M 0
• Nonlinear control laws have advantages
– f has vertical asymptote: bounded error e
– f has horizontal asymptote: bounded effort
u
– Possible to converge in finite time.
– Nonlinearity allows more kinds of
composition.
 Stopping Controller
• Desired stopping point: x=0.
– Current position: x
d = | x | +ε
– Distance to obstacle:

v = x˙ = − f (x)
• Simple P-controller:

f (x) =for
• Finite stopping time k | x | sgn(x)
 Derivative Control
• Damping friction is a force
opposing motion, proportional to
velocity.
• Try to prevent overshoot by damping
controller response.
u = −kP e − kD e˙
• Estimating a derivative from
measurements is fragile, and
amplifies noise.
Derivative Control in
Action

– Damping fights oscillation and overshoot

– But it’s vulnerable to noise
Effect of Derivative
Control

– Different amounts of damping (without noise)

 Derivatives Amplify
Noise

– This is a problem if control output

(CO) depends on slope (with a high
gain).
 The PID Controller
• A weighted combination of
Proportional, Integral, and Derivative
terms. t
u(t) = −kP e(t) − kI ∫ e dt − k
D e˙(t)
0
• The PID controller is the workhorse of
the control industry. Tuning is non-
trivial.
– Next lecture includes some tuning methods.
PID Control in Action

– But, good behavior depends on good tuning!

– Aibo joints use PID control
Exploring PI Control
Tuning
 Habituation
• Integral control adapts the
bias term ub.
• Habituation adapts the
setpoint xset.
– It prevents situations where too
much control action would be
u = −kP e + ub
dangerous.
• Both adaptations reduce
x˙ set = +k h e where k h << k P
steady-state error.
 Types of Controllers
• Open-loop control
– No sensing
• Feedback control (closed-loop)
– Sense error, determine control response.
• Feedforward control (closed-loop)
– Sense disturbance, predict resulting error,
respond to predicted error before it
happens.
• Model-predictive control (closed-loop)
– Plan trajectory to reach goal.
– Take first step.
– Repeat. Design open and closed-loo
controllers for me to get
of the room.
 Dynamical Systems
• A dynamical system changes
continuously (almost always)
n
x˙ = F(x) to
according where x ∈ ℜ

• A controller is defined to
change the coupled robot and
x˙ = F(x,u) into a desired
environment
˙
x
dynamical system. = F(x,H i (G(x)))
y = G(x)
u = H i (y) x˙ = Φ(x)
Two views of dynamic
behavior
• Time
plot
(t,x)

• Phase
portrai
t (x,v)
 Phase Portrait: (x,v)
space
• Shows the trajectory (x(t),v(t)) of the system
– Stable attractor here
 In One Dimension
• Simple linear
x˙
x˙ = kx
system

• Fixed point
x =0 ⇒ x˙ = 0 x
• Solution
kt
x(t) = x 0 e
– Stable if k < 0
– Unstable if k > 0
 In Two Dimensions
• Often, we have position and
T
velocity:
x = (x,v) where v = x˙

• If we model actions as
forces, which cause
⎛x˙ ⎞ ⎛x˙ ⎞ ⎛ v ⎞
x˙ = ⎜ ⎟ = ⎜ ⎟ =then
acceleration,
⎜ ˙ ⎟ ⎜˙˙⎟ ⎜ ⎜ we⎟get:
⎟
⎝v ⎠ ⎝x ⎠ ⎝ forces⎠
 The Damped Spring
• The spring is defined by Hooke’s
Law: F = ma = m&x&= −k x 1

x&= −k x friction
&
• Includemdamping
1−k x& 2

x&+and
&
• Rearrange bx&+redefine
cx= 0 constants
⎛x˙ ⎞ ⎛x˙ ⎞ ⎛ v ⎞
x˙ = ⎜⎜ ⎟⎟ = ⎜⎜ ⎟⎟ = ⎜⎜ ⎟⎟
⎝v˙ ⎠ ⎝x˙˙⎠ ⎝−bx˙ − cx ⎠
 Node
Behavi
or
 Focus
Behavi
or
Saddle
Behavi
or
Spiral
Behavi
or

(stabl
e
attrac
tor)
Center
Behavio
r

(undamp
ed
oscilla
tor)
The Wall Follower

(x,
y) θ
 The Wall Follower
• Our robot model:
⎛x˙ ⎞ ⎛v cosθ ⎞
⎜ ⎟ ⎜ ⎟
x˙ = ⎜y˙ ⎟ = F (x,u) = ⎜v sinθ ⎟
⎜˙⎟ ⎜ ⎟
⎝θ ⎠ ⎝ ω ⎠

u = (v )T y=(y )T

  0.

• We set the control law u = (v

)T = Hi(y)
 The Wall Follower
• Assume constant forward velocity v =
v0
– approximately parallel to the wall: 
 0.
e = y − ydistance
• Desired set
˙ = y˙ wall
so efrom e˙˙ = y˙˙
and defines
error:

• We set the control law u = (v )T

= Hi(y) ˙e˙ + k e˙ + k e = 0
1 2
– We want e to act like a “damped spring”
 The Wall Follower
˙e˙ + k1 e˙ +
• We want a damped spring: k2 e = 0
• For small values of 
e˙ = y˙ = v sinθ ≈ vθ
˙e˙ = ˙y˙ = v cosθ θ˙ ≈ vω
• Substitute, and assume v=v0 is
constant.
v 0 ω + k1 v 0 θ + k 2 e = 0

• Solve for 
 The Wall Follower
˙e˙ + k1 e˙ + k2 e = 0
• To get the damped spring
• We get the constraint
v 0 ω + k1 v 0 θ + k 2 e = 0
• Solve for . Plug into u.
⎛ v ⎞
⎛v ⎞ ⎜ 0

€ u = ⎜ ⎟ = −k θ − k 2 e⎟ = H i (e,θ)
⎝ω ⎠ ⎜ 1 v ⎟
⎝ 0 ⎠
– This makes the wall-follower a PD
controller.
– Because:
€
 Tuning the Wall
Follower
• The system˙e˙ is
+ k1 e˙ + k2 e = 0
2
• Critical damping requires k1 − 4k 2 = 0

k1 = 4k 2
• Slightly underdamped performs
better.
4k2
– Set k2 by experience.
– Set k1 a bit less than
 An Observer for
Distance to Wall
• Short sonar returns are
reliable.
– They are likely to be
perpendicular reflections.
 Alternatives
• The wall follower is a PD control
law.
• A target seeker should probably be a
PI control law, to adapt to motion.

• Can try different tuning values for

parameters.
– This is a simple model.
– Unmodeled effects might be significant.
Ziegler-Nichols Tuning
• Open-loop response to a unit
step increase.
• d is deadtime. T is the process
time constant.
• K is the process gain.

d T K
 Tuning the PID
Controller
• We have described it as:
t
u(t) = −kP e(t) − kI ∫ e dt − k
D e˙(t)
0
• Another standard form is:
⎡ t ⎤
u(t) = −P ⎢e(t) + TI ∫ e dt + TD e˙(t)⎥
⎣ 0 ⎦
• Ziegler-Nichols says:
1.5 ⋅T
P= TI = 2.5 ⋅ d TD = 0.4 ⋅ d
€ K ⋅d
Ziegler-Nichols Closed
Loop
1. Disable D and I action (pure P
control).
2. Make a step change to the
setpoint.
3. Repeat, adjusting controller
gain until achieving a stable
oscillation.
• This gain is the “ultimate gain” Ku.
• The period is the “ultimate period”
Pu.
 Closed-Loop Z-N PID
Tuning
• A standard form of PID is:
⎡ t ⎤
u(t) = −P ⎢e(t) + TI ∫ e dt + TD e˙(t)⎥
⎣ 0 ⎦
• For a PI controller:
Pu
P = 0.45 ⋅K u TI =
1.2
€
• For a PID controller:
Pu Pu
€ P = 0.6 ⋅K u TI = TD =
2 8
 Summary of Concepts
• Dynamical systems and phase portraits
• Qualitative types of behavior
– Stable vs unstable; nodal vs saddle vs
spiral
– Boundary values of parameters
• Designing the wall-following control
law
• Tuning the PI, PD, or PID controller
– Ziegler-Nichols tuning rules
– For more, Google: controller tuning
 Followers

• A follower is a control law

where the robot moves forward
while keeping some error term
small.
– Open-space follower
– Wall follower
– Coastal navigator
– Color follower
 Control Laws Have
Conditions
• Each control law includes:
– A trigger: Is this law
applicable?
– The law itself: u = Hi(y)
– A termination condition: Should
the law stop?
 Open-Space Follower
• Move in the direction of large
amounts of open space.
• Wiggle as needed to avoid
specular reflections.
• Turn away from obstacles.
• Turn or back out of blind
alleys.
 Wall Follower
• Detect and follow right or
left wall.
• PD control law.
• Tune to avoid large
oscillations.
• Terminate on obstacle or wall
vanishing.
 Coastal Navigator
• Join wall-followers to follow a
complex “coastline”
• When a wall-follower terminates,
make the appropriate turn, detect a
new wall, and continue.
• Inside and outside corners, 90 and
180 deg.
• Orbit a box, a simple room, or the
desks.
 Color Follower
• Move to keep a desired color
centered in the camera image.
• Train a color region from a
given image.
• Follow an orange ball on a
string, or a brightly-colored
T-shirt.
Problems and Solutions
• Time delay
• Static friction
• Pulse-width modulation
• Integrator wind-up
• Chattering
• Saturation, dead-zones,
backlash
• Parameter drift
 Unmodeled Effects
• Every controller depends on its
simplified model of the world.
– Every model omits almost everything.
• If unmodeled effects become
significant, the controller’s model
is wrong,
– so its actions could be seriously wrong.
• Most controllers need special case
checks.
– Sometimes it needs a more sophisticated
model.
 Time Delay
t1 t t2

now
• At time t,
– Sensor data tells us about the world at
t1 < t.
– Motor commands take effect at time t2 >
t.
– The lag is dt = t2  t1.
• To compensate for lag time,
– Predict future sensor value at t2.
– Specify motor command for time t2.
 Predicting Future
Sensor Values
• Later, observers will help us make
better predictions.
• Now, use a simple prediction
method:
– If sensor s is changing at rate ds/dt,
– At time t, we get s(t1), where t1 < t,
– Estimate s(t2) = s(t1) + ds/dt * (t2 -
t1).
• Use s(t2) to determine motor signal
u(t) that will take effect at t2.
 Static Friction
(“Stiction”)
• Friction forces oppose the direction of
motion.
• We’ve seen damping friction: Fd =  f(v)
• Coulomb (“sliding”) friction is a
constant Fc depending on force against
the surface.
– When there is motion, Fc = 
– When there is no motion, Fc =  + 
• Extra force is needed to unstick an
object and get motion started.
 Why is Stiction Bad?
• Non-zero steady-state error.
• Stalled motors draw high current.
– Running motor converts current to
motion.
– Stalled motor converts more current to
heat.
• Whining from pulse-width
modulation.
– Mechanical parts bending at pulse
frequency.
Pulse-Width Modulation
• A digital system works at 0 and 5
volts.
– Analog systems want to output control
signals over a continuous range.
– How can we do it?
• Switch very fast between 0 and 5 volts.
– Control the average voltage over time.
• Pulse-width ratio = ton/tperiod. (30-50
sec)

ton

tperiod
 Pulse-Code Modulated
Signal
• Some devices are controlled by the
length of a pulse-code signal.
– Position servo-motors, for example.

0.7ms

20ms

1.7ms

20ms
 Integrator Wind-Up
• Suppose we have a PI controller
t
u(t) = −kP e(t) − k I ∫ e dt + u
b
0
• Motion might be blocked, but the
integral is winding up more and more
control action.

€ u(t) = −kP e(t) + ub

u˙ b (t) = −kI e(t)
• Reset the integrator on significant
events.
 Chattering
• Changing modes rapidly and
continually.

– Bang-Bang controller with thresholds

set too close to each other.

– Integrator wind-up due to stiction

near the setpoint, causing jerk,
overshoot, and repeat.
 Dead Zone
• A region where controller output
does not affect the state of the
system.
– A system caught by static friction.
– Cart-pole system when the pendulum is
horizontal.
– Cruise control when the car is stopped.

• Integral control and dead zones can

combine to cause integrator wind-up
problems.
 Saturation
• Control actions cannot grow
indefinitely.
– There is a maximum possible output.
– Physical systems are necessarily
nonlinear.

• It might be nice to have bounded

error by having infinite response.
– But it doesn’t happen in the real world.
 Backlash
• Real gears are not perfect
connections.
– There is space between the teeth.

• On reversing direction, there

is a short time when the input
gear is turning, but the
output gear is not.
 Parameter Drift
• Hidden parameters can change the
behavior of the robot, for no obvious
reason.
– Performance depends on battery voltage.
– Repeated discharge/charge cycles age the
battery.
• A controller may compensate for small
parameter drift until it passes a
threshold.
– Then a problem suddenly appears.
– Controlled systems make problems harder to
find
 Unmodeled Effects
• Every controller depends on its
simplified model of the world.
– Every model omits almost everything.
• If unmodeled effects become
significant, the controller’s model
is wrong,
– so its actions could be seriously wrong.
• Most controllers need special case
checks.
– Sometimes it needs a more sophisticated
model.