Collaborative Teleoperation via the Internet Ken Goldberg, Billy Chen, Rory Solomon, Bob Farzin, Derek Poon, Jacob Heitler, Steve Bui UC Berkeley.

Collaborative Teleoperation via the

Internet

Ken Goldberg, Billy Chen, Rory Solomon, Bob Farzin, Derek Poon, Jacob Heitler,

Steve Bui UC Berkeley

Jason: an underwater robot

Collaborative Robot Control:

Camera

Camera

Ca

me

ra

Light

Jason

Related Work R Goertz and R. Thompson, 1954. H. Linstone and M. Turoff. 1975. A. K. Bejczy. 1980. R. D. Ballard, 1986. T. Sheridan. 1992. R. and L. Carpenter. 1993. O. Khatib, et al. 1996. K. Goldberg, et al. 1995. E Paulos and J. Canny. 1996. B. Dalton and K. Taylor. 1998. Y. Akatsuka, S. Goldberg, G, Bekey. 1998. R. Siegwart. 1998.

Delphi Method Developed at RAND Corporation in 1960’s a method of iterated polling a group submits votes, views results, discusses

results and repeats experimental results have shown that this

method converges very closely to the correct result when the group is posed with a question

Cinematrix

• Audience Members work collaboratively

• Everyone holds a red/green “paddle”

How can multiple users control one robot?

?…

User 1User 2

User n

The resource (robot) must use a control model which satisfies each user in a fair way.

Batch Model

In the 1994 Mercury project, the users queued up for 5 minute individual control sessions.

…

User 1

User 2

User n

User n…

User 1User 2

Multitask

In the 1995 Telegarden project, user motion requests were interleaved – simulating simultaneous control.

…

User 1User 2

User n time

Central Limit Theorem

Consider n noisy iid inputs: Xi.

As the number of inputs increases, the noise in the aggregate decreases with the square root of n.

n

X* =Xi ) - µ(

Collaborative Control aggregates user inputs:

In our control model, we aggregate all the user requests into a single request (collaborative control).

…

User 1User 2

User n

Aggregate

…

User 2

User n

…

User 1User 2

User n

…

User 1

…User 2

User n Collaborative

Control methods:

MultiTask

Batch

Application: Ouija board

A classic “game” One “planchette” Many players

System Architecture

Client

Internet

Webserver Video Card Robot Control Webserver C

CCD Camera Robot

Applet V Applet C

Client Interface: Java applet

• Top Window: video applet

• Streaming live video of the robot

• Bottom Window: planchette applet

Planchette Motion Model

Realistic motion Inertia Frictional drag

Hardware

• Two PC servers: one running Win98, one running Linux• Adept 604-S Robot arm • Panasonic CCD camera

Challenges Serial connetion OS: Linux vs. Windows streaming video: server-push vs.

Java applet: Inetcam

To test collaborative performance we used a maze.

Views with Noise:

Performance Tests Time to complete maze (sec):

Subject Trial 1 Trial 2 Trial 3A 146 139 105B 177 141 175

A,B 65 71 72

New users per day:

0

50

100

150

200

250

300

350

12/9/

99

12/23

/99

1/6/

00

1/20

/00

2/3/

00

2/17

/00

3/2/

00

3/16

/00

3/30

/00

4/13

/00

Currently 8000 registered players

Summary Introduction

Coping with many noisy sensors Batch, Multitasking, and Collaborative

control models Implementation

User Interface and Motion Model Experimental Results and

Conclusion

Future Work Different control models

Winner-take-all Input weighting based on prior

performance Different test conditions

Different time delays Different goals

http://ouija.berkeley.edu