Cyber Physical Computing Systems

Mobile Robots

Pangun Park

Chungnam National University

[email protected]

Pangun Park (CNU) 1 / 29

Driving Robots Around?

What does it take to drive a robot from point A to point B?

controller

sensors
robot model

Pangun Park (CNU) 2 / 29

Divide and Conquer

The world is dynamic and fundamentally unknown

The controller must be able to respond to environmental conditions

Instead of building one complicated controller divide and conquer: Behaviors

Go-to-goal
Avoid-obstacles
Follow-wall
Track-target

Pangun Park (CNU) 3 / 29

Behaviors

Pangun Park (CNU) 4 / 29

Differential Drive Robots

In order to control mobile robots, we need models

Differential drive wheeled robots a very common type

Pangun Park (CNU) 5 / 29

Model 1.0

vℓ vr
R
φ

(x, y)
L  R
 ẋ = (v + v ) cos φ

R
ẋ
 = 2
(vr + vl ) cos φ
R
ẏ = 2
(vr + vl ) sin φ
R

φ̇ = (v − vl )

L r

Pangun Park (CNU) 6 / 29

The “Unicycle” Model

But it is not very natural to “think” in terms of wheel velocities

Go directly for translational and angular velocities

Inputs:

v
ω

Dynamics:
φ 
ẋ
 = v cos φ
ẏ = v sin φ
(x, y) 
φ̇ =ω


Pangun Park (CNU) 7 / 29

Model 2.0

Design for this model!


ẋ
 = v cos φ R 2v
v= (vr + vl ) → = vr + vl
ẏ = v sin φ 2 R
R ωL

φ̇ =ω

ω = (vr − vl ) → = vr − vl
L R

Implement this model!


R
ẋ
 = 2
(vr + vl ) cos φ 2v + ωL
R vr =
ẏ = 2
(vr + vl ) sin φ 2R
R 2v − ωL

φ̇ = (vr − vl )

L vl =
2R

Pangun Park (CNU) 8 / 29

Odometry

The state of the robot is (x, y , φ)

How do we obtain this state information?

Two possibilities:
External sensors
Internal sensors
Orientation: Compass, ...
Position: Accelerometers, Gyroscopes, ...
Wheel Encoders

Pangun Park (CNU) 9 / 29

Wheel Encoders

Wheel encoders give the distance moved by each wheel

Assume the wheels are following an arc (short time scale)

Dr

Dc Dl + Dr
Dc =
2
Dℓ


ẋ
 = x + Dc cos φ
ẏ = y + Dc sin φ
= φ + Dr −D

l
φ̇

L

Pangun Park (CNU) 10 / 29

Wheel Encoders

But how do we know how far each wheel has moved?

Dr • Assume each wheel has N “ticks”

per revolution
Dc
• Most wheel encoders give the total
Dℓ
tick count since the beginning

For both wheels:

∆tick = tick′ − tick

∆tick
D = 2πR
N

Pangun Park (CNU) 11 / 29

A Major Disclaimer

DRIFT!!!

Pangun Park (CNU) 12 / 29

Sensors

Robots need to know what the world around them looks like

The standard sensor suite includes a “skirt” of range sensors:

IR, Ultra-Sound, LIDAR, ...

Other standard external sensors include

Vision
Tactile
“GPS”

Pangun Park (CNU) 13 / 29

Range-Sensor Skirts

We will mainly deal with range-sensors

Pangun Park (CNU) 14 / 29

The Disk Abstraction

Instead of worrying about the resolution of the sensors, assume we know the distance and
direction to all obstacles around us (that are close enough)

(d2 , φ2 ) Angles are relative to

the robot’s heading

(d1 , φ1 )

Pangun Park (CNU) 15 / 29

The Disk Abstraction

Instead of worrying about the resolution of the sensors, assume we know the distance and
direction to all obstacles around us (that are close enough)

• If we know our own poseknow our own pose (position

If we
(d2 , φ2 ) (position and orientation):
and orientation):

(d1 , φ1 )
x1 = x + d1 cos(φ1 + φ)
y1 = y + d1 sin(φ1 + φ)

Pangun Park (CNU) 16 / 29

Example: Rendezvous

Pangun Park (CNU) 17 / 29

Behavior-Based Robotics

The world is fundamentally

unknown and changing

Does not make sense to over-plan

Key idea: Develop a library of

useful controllers (=behaviors)

Switch among controllers in

response to environmental changes

Pangun Park (CNU) 18 / 29

Building a Behavior v.1

Assume we have a differential-drive, wheeled mobile robot driving at a constant speed


ẋ = v0 cos φ

ẏ = v0 sin φ

φ̇ = ω


φd • Want to drive in a desired heading

φ ω =???

Pangun Park (CNU) 19 / 29

Building a Behavior v.1

We have a reference, a model, a control input, and a tracking error:

r = φd , e = φd − φ, φ̇ = ω

Why not use PID?

Z
ω = kP e + kI edτ + kD ė

Pangun Park (CNU) 20 / 29

Dealing with Angles

This typically will not work since we are dealing with angles:

φd = 0, φ = 100π → e = −100π

Solution: Ensure that e ∈ [−π, π]

Standard trick is to use atan2

e ′ = atan2(sin(e), cos(e)) ∈ [−π, π]

Pangun Park (CNU) 21 / 29

Example: Navigation

Problem: Go to a goal location without bumping in to obstacles:

obstacles:

At a minimum, we need two behaviors: Go-to-goal and Avoid- obstacles

Pangun Park (CNU) 22 / 29

Go-To-Goal

How drive a robot to a goal location?


ẋ = v0 cos φ

ẏ = v0 sin φ
 e = φd − φ, ω = PID(e)
φ̇ = ω


(xg , yg )
! "
yg − y
φd = arctan
xg − x
(x, y)

Pangun Park (CNU) 23 / 29

Attempt 1

ω = K (φd − φ)

ANGLES!!!

Pangun Park (CNU) 24 / 29

Attempt 2

“ω = K (φd − φ)”

GAIN TOO LOW!!!

Pangun Park (CNU) 25 / 29

Attempt 3

“ω = Kbig (φd − φ)”

big

JUST RIGHT!!!

Pangun Park (CNU) 26 / 29

Obstacle-Avoidance

How avoid driving into obstacles?

We can use the same idea by defining a desired heading

(xo , yo )

(x, y) There is no obvious

“correct” answer

Pangun Park (CNU) 27 / 29

Where to Go?

Choice depends on direction to goal.

Not “pure” but “blended”

π
φ2 = φobst ±
2
φ3 = φgoal
Pure go-to-goal
φ4 = F (φobst , φgoal )
Blended
φ1 = φobst + π
This is “pure” avoidance

Pangun Park (CNU) 28 / 29

Arbitration Mechanisms

This example illustrate two fundamentally different “arbitration mechanisms”

Winner takes all = Hard switches
Blending = Combined behaviors

Both approaches have merit in different situations

Performance?
Analysis?

We will see how to design systematic behaviors and arbitration mechanisms

Pangun Park (CNU) 29 / 29