Vision Based Motion Control Martin Jagersand University of Alberta CIRA 2001

Vision Based Motion Control

Martin Jagersand

University of Alberta

CIRA 2001

Vision Based Motion Control

Martin Jagersand

University of Alberta

CIRA 2001

Content

1. Vision based motion control

2. Programming and solving whole human tasks

3. Software systems for vision and control

4. Discussion

1. How to go from Visual sensationto Motor action?>

Camera -> Robot coord Robot -> Object

Closed loop traditional visual servoing

This talk: focus on estimating the geometric transforms

EE

Lots of possible coordinates

Camera– Frame at projection center– Many different models

Robot– Base frame– End-effector frame– Object frame

Traditional modeling: P=P1(<params>) P2(<params>)… Pn(<params>)

Hand-Eye system

Motor-Visual function: y=f(x)Jacobian: J=( dfi / dxj )

Recall:Visual specifications

Point to Point task “error”:

yãE = yã à y0 y0

yã

y0

E =y1...y16

2

4

3

5

ã

ày1...y16

2

4

3

5

0

Why 16 elements?

Visual Servoing

Observed features: Motor variables: Local linear model: Visual servoing steps: 1 Solve:

2 Move:

y = y1 y2 . . . ym[ ]T

x = x1 x2. . . xn[ ]T

É y = J É x

yã à yk = J É x

xk+1 = xk + É x

Find J Method 1: Test movements along basis

Remember: J is unknown m by n matrix

Assume movements Finite difference:

J =@x1

@f1 ááá @xn

@f1

.... . .

@x1

@fm

@xn

@fm

0

B@

1

CA

É x1 = [1;0; . . .;0]T

É x2 = [0;1; . . .;0]T...É xn = [0;0; . . .;1]T

J t

...É y1...

2

4

3

5

...É y2...

2

4

3

5 ááá

...É yn...

2

4

3

5

0

@

1

A

Find J Method 2:Secant Constraints

Constraint along a line: Defines m equations Collect n arbitrary, but different measures y Solve for J

É y = J É x

ááá É yT1 ááá

â ã

ááá É yT2 ááá

â ã...

ááá É yTn ááá

â ã

0

BB@

1

CCA =

ááá É xT1 ááá

â ã

ááá É xT2 ááá

â ã...

ááá É xTn ááá

â ã

0

BB@

1

CCA J T

Find J Method 3:Recursive Secant Constraints

Based on initial J and one measure pair Adjust J s.t. Rank 1 update:

Consider rotated coordinates: – Update same as finite difference for n orthogonal

moves

É y;É x

É y = J k+1É x

Jêk+1 = Jêk +É xTÉ x

(É y à JêkÉ x)É xT

Trust region of J estimate

Let be the trust region at time t Define a model agreement:

Update the trust region recursively:

dt = É ymeasuredk kJêÉ xk k

ët+1 =21ët if dt ô dlower

ët if dlower < dt ô dupper

max(2É xt;ët) if dt > dupper

8<

:

Where dupper and are dlower predefined constants

ët

Visual Servoing Steps

1. Solve:

2. Update and move:

3. Read actual visual move

4. Update Jacobian:

yã à yk = J É x

xk+1 = xk + É x

Jêk+1 = Jêk +É xTÉ x

(É ym à JêkÉ x)É xTÉ ym

Visual Servoing Steps

1. Solve:

2. Update and move:

3. Read actual visual move

4. Update Jacobian:

yã à yk = J É x

xk+1 = xk + É x

Jêk+1 = Jêk +É xTÉ x

(É y à JêkÉ x)É xT

Jacobians = Spline model of underlying non-linear function

Over time acquires several Jacobians J

Each J a hyperplane Collection of J’s form

a (sparse) piecewise linear spline

Jacobian based visual model

Assume visual features m>>n motor freedoms

All visual change restricted to n freedoms by:

1. Can predict visual change

2. Can also parameterize x visually

É y = J É x

Related visual model:Affine model

Affine basis

Image projection of origin:

Image basis:

(e0; . . .;e3) = ((0;0;0);(0;0;1); . . .; (1;0;0))

P0(y) = y1

y2

ò ó

P1(y) = y3 y5 y7

y4 y6 y8

ò óà P0(y) 1 1 1( )

ô õ:

e1

e2

e3

O

Find affine coordinates

Observe (track) y through time Solve an equation system to find q

Reprojection: Have q,want y

p = P1(y)q + P0(y) = P(y;q)

p1 à P0(y1)...pw à P0(yw)

0

B@

1

CA =

P1(y1)...P1(yw)

0

B@

1

CA q

e1

e2

e3

Oq

Relation Affine – Jacobian image models

Rewrite affine model

p1x...p

2xm

p1y...p

2ym

0

BBBBBB@

1

CCCCCCA

=

qT1... 0

qT2m

qT1

0...

qT2m

0

BBBBBBB@

1

CCCCCCCA

y3 à y1

y5 à y1

y7 à y1

y4 à y2

y6 à y2

y8 à y2

0

BBBBBBB@

1

CCCCCCCA

+

y1...y1

y2...y2

0

BBBBB@

1

CCCCCA

Composite affine and Jacobian model

Chain the affine and Jacobian model Represents rigid objects in arbitrary motor

frame

p1x p1y...p

2xm p

2ym

0

@

1

A =1...1

!

y1 + J 0x; y2 + J 1x( ) +

+qT

1...qT

2m

0

@

1

A J 2::4x; J 5::7x( ) +y3; y4

y5; y6

y7; y8

0

@

1

A

0

@

1

A

Transforms Affine-Jacobian model

Measurement matrix

Affine coordinate equation:

M = M b

M d

ò ó= J tX + f (xt)(1; . . .;1)

M Td à

P0(y1)...P0(yw)

0

B@

1

CA (1;. . .;1) =

P1(y1)...P1(yw)

0

B@

1

CA Q

Experiment:Affine animation of rigid structure

Affine vs. Visual-Motor

Other sensory modalities: Force and contact manipulation

Accuracy is limited by: Visual tracking

and Visual goal specification Specifying well defined visual encodings can

be difficult Limited to non-occluded settings Not all tasks lend themselves to visual

specification.

Constraint Geometry

Impact force along surface normal:

Sliding motion:

3rd vector:

p1 = jfkjfk

p2 = jî xkjî xk

p3 = p1 â p2 = jfkjfk â

jî xkjî xk

Constraint Frame

With force frame = tool frame we get:

Assume frictionless => Can update each time step

Pk+1 =p1p2p3

0

@

1

A =

jfkjfk

jî xkjî xk

jfkjfk â

jî xkjî xk

0

BBB@

1

CCCA

P1

P2

P3

Hybrid Control Law

Let Q Joint -> Tool Jacobian Let S be a switching matrix, e.g. diag([0,1,1]) Velocity control u:

uk = à K vQà 1Pà 1k

SPkQJ à 1ek à K cQà 1Pà 1k

(I à S)Pkfk

Visual part Force part

Accounting for Friction

Friction force is along motion direction! Subtract out to recover surface normal:

Pk+1 =

jfkà pT

2fkp2j

fkà pT

2fkp2

jî xkjî xk

jfkà pT

2fkp2j

fkà pT

2fkp2

âjî xkjî xk

0

BBBBBB@

1

CCCCCCA

Motion Sequence

Motion Sequence

Summary of model estimation and visual motion control

Model estimation is on-line and requires no special calibration movements

Resulting Jacobians both model/constrain the visual situation and provide visual motor transf.

Motion control is direct from image based error functions to motor control. No 3D world space.

2. How to specify a visual task sequence?

1. Grasp

2. Move in

3. Cut

1. Grasp

2. Reach close

3. Align

4. Turn

Recall: Parallel Composition Example:

E (y ) =wrenchy - y4 7

y - y2 5

y • (y y )8 3 4

y • (y y )6 1 2

Visual error function “spelled out”:

Serial CompositionSolving whole real tasks

Task primitive/”link”

1. Acceptable initial (visual) conditions

2. Visual or Motor constraints to be maintained

3. Final desired condition Task =

A = (E init;M;E final)

A1A2...Ak

“Natural” primitive links

1. Transportation Coarse primitive for large movements <= 3DOF control of object centroid Robust to disturbances

2. Fine Manipulation– For high precision control of both position and

orientation– 6DOF control based on several object features

Example: Pick and place type of movement

3. Alignment??? To match transport final to fine manipulation initial conditions

More primitives

4. Guarded move– Move along some direction until an external

contraint (e.g. contact) is satisfied.

5. Open loop movements: When object is obscured Or ballistic fast movements Note can be done based on previously estimated

Jacobians

Solving the puzzle…

Teaching and Programming in Visual Space

1. Tele Assistance A tele-operator views the scene through stereo cameras Objects to be manipulated are pointed out on-line

2. Visual Programming Off-line Like xfig, macpaint, but with a palette of motor actions.

3. Teaching by Showing A (human) manipulation is tracked in visual space The tracked data is used to (automatically?) generate a

sequence of visual goals

HCI: Direct manipulationExample: xfig drawing program

Icons afford use Results visible Direct spatial action-

result mapping

line([10, 20],[30, 85]);patch([35, 22],[15, 35], C);

% C complex structuretext(70,30,'Kalle'); % Potentially add font, size, etc

matlab drawing:matlab drawing:

Example:Visual programming

Task control summary

Servoing alone does not solve whole tasks– Parallel composition: Stacking of visual constraints

to be simultaneously satisfied– Serial composition: Linking together several small

movements into a chain of continuous movements

Vision-based user interface– Tele-assistance– Visual Programming– Teach by showing

Types of robotic systems

Autonomy

Generality

Supervisory control

Tele-assistance

Programming by demonstration

Preprogrammed systems

3. Software systems for vision-based control

Hand-EyeSystem

System requirements

Solve many very different motion tasks– Flexible, teachable/re-programmable

Real time– On special embedded computers or general

workstations

Different special HW Multiprocessors

Toolbox

System design

Interpreted “scripting” language gives flexibility Compiled language needed for speed and HW

interface.

Examples Matlab

Greencard

Haskell

C, C++, fortran

PVM

C, C++

Dyn linking (mex)

Usage example:

Specialize robot– projandwait(zero3,’robotmovehill’,A3D,’WaitForHill’);

Initialize goals and trackers– [TrackCmd3D,N] = InitTrackers([1 1],[0,1]);– PU = GetGoals([1 1],[0,1]);

Servo control– J3s = LineMove(‘projandwait’,TrackCmd3D,J3i,PU,Ndi,err)

Software systems summary

Most current demos solve one specific movement

For solving many everyday tasks we need flexibility and reprogrammability– Compiled primitive visual trackng and– Interpreted scripting language– Higher order functions

Workshop conclusions

?

Workshop conclusions

Sensing is unreliable and incomplete– Can’t reliably build internal 3D world models, but can

use the real world as an external reference.

A-priori object and world models uncommon in human environments– Estimate on-line and only what’s needed.

Human users require human interaction techniques– Interacting by visual pointing and gestures is natural.

Action/Perception division in human and machine hand-eye syst.

Open questions?

For particular tasks what are the most natural representations and frames?

Global convergence of arbitrarily composed visual error functions?

Robustness? Interaction with other sensing modalities?

Feedback system

Fast internal feedback Slower external trajectory corrections

Short and long control loops

Applications for vision in User Interfaces

Interaction with machines and robots– Service robotics– Surgical robots– Emergency response

Interaction with software– A store or museum information kiosk

Service robots

Mobile manipulators, semi-autonomous

DIST TU Berlin KAIST

TORSO with 2 WAMs

Service tasks

This is completely hardwired! Found no real task on WWW

But

Maybe first applications in tasks humans can’t do?

Why is humanlike robotics so hard to achieve?

See human task:– Tracking motion, seeing gestures

Understand:– Motion understanding: Translate to correct

reference frame– High level task understanding?

Do: – Vision based control