Video of our “Regis” prototype running a demonstration we showed at AAAI-07 in Vancouver. For more information, see the Regis page.


MPEG-4 High-res (.mp4, 12.7MB)
MPEG-4 Low-res (.mp4, 7.5MB)
YouTube (Flash)

Generated by streaming video over a wireless connection to a PC, where the individual frames are saved to disk and then compressed into video format.

Raw: 625 KB, 16 seconds (40 KB/s)
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Sony Aibo ERS-210A)

Segmented: 1.2 MB, 16 seconds (81 KB/s)
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Sony Aibo ERS-210A)

• High quality compression was used for the segmented video to retain its crisp edges.
• Segmented video is full resolution of what the Aibo is processing.  Raw video is half resolution of what the Aibo is "seeing".
• See the screenshots for the Raw GUI and Segmented GUI that produced these files.
• Other available GUI tools are listed on the TekkotsuMon Tutorial page.

The Aibo plays on a laptop simulating a gambling machine. This utilizes state machine, force feedback, and vision:

4.3 MB, 70 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Minolta DiMage 7i, edited in iMovie)

• Major objects used: BanditMachine (source: .h)
• You'll also need some code to run on the laptop:
  • Mac OS X: App (source: .sit)
  • Java: App (source: .java)

Showcases the walking abilities included with the framework.  The walking algorithms are ported from CMU's RoboSoccer team.  The low walking style is both fast and allows the ball to be dribbled. It also lowers the center of gravity, which makes it harder to knock over.  Notice how the gait transitions smoothly, without stumbling.  However, it is optimized for level, even terrain, and may not perform as well elsewhere.

5.4 MB, 90 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
-  AVI (.avi)

(shot on a Minolta DiMage 7i, edited in iMovie)

3.1 MB, 51 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
-  AVI (.avi)

(shot on a Minolta DiMage 7i, edited in iMovie)

• Major objects used:
  • FollowHeadBehavior (source: .h/.cc)
  • WalkMC (source: .h/.cc)

A project by one of the Pennsylvania Governor's School for the Sciences 2004 team projects, this video shows the AIBO centering between walls, aligning to the walls, and exploring a maze.  Final report is available, The Sony AIBO: Using IR for Maze Navigation, by Kyle Lawton and Elizabeth Shrecengost, advised by Ethan Tira-Thompson and Zhan Shi.

3.8 MB, 62 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Canon S410)

• Major sources used (not included in releases, available from CVS contrib directory)
  
  • MazeRunner.h
  • firstAlign.h
  • Alignment.h
  • Decide.h

Uses the Aibo's accelerometers to keep the head in a constant orientation, regardless of the angle the body. Pressing the head buttons releases the neck joints and allows the head to be positioned where it should be held.

2.5 MB, 41 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Minolta DiMage 7i)

• Major objects used:
  • HeadLevelBehavior (source: .h)
  • HeadPointerMC (source: .h/.cc)

Using Y-channel (intensity) images streamed over wireless and recorded on a PC.  Post processing in MATLAB allows rapid test and development of vision algorithms.  In this case, we try to stabilize the camera in software by registering each new image to the one before it.  Unfortunately, the computational cost of this example appears to be too high to actually implement on the AIBO itself. (Might be possible, but difficult)  The current walking target velocity is also logged and used to predict "intentional" rotational displacement between images.

3.2 MB, 43 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)


(shot on a Sony Aibo ERS-210A)

82 seconds
- MPEG-4 (.mp4, 6.2 MB)
- Quicktime (.mov, 5.2 MB)


(shot on a Sony Aibo ERS-210A)

• Class presentation is here (PDF) and here (PPT)
• Code is available from here (.tgz)

Demonstrates the default startup mode.  A simple double-tap on the back button at any time will enable or disable this mode.  The tail light pulses blue/red to signal this mode, and the joints remain pliable, for both safety and allowing the robot to be posed manually.

This example shows the ERS-210, but the ERS-220 and ERS-7 all use unique LEDs, so they have different signaling patterns.

2.5 MB, 41 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot on a Minolta DiMage 7i, edited in iMovie)

• Major objects used:
  • EmergencyStopMC (source: .h/.cc)

The Aibo learns that a single ball predicts reward, but both balls predict no reward.  This uses our internal AI module (under development), object recognition, and (simple) gesture recognition.

4.0 MB, 65 seconds
- MPEG-4 (.mp4)
- Quicktime (.mov)
- AVI (.avi)

(shot  on a DV camera, edited in iMovie)
The paper that describes the research aspects of this demo can be found here.

The following screenshots are from the Java and MATLAB tools which are bundled with Tekkotsu.  Click on a name to get a description of the tool, or click on the thumbnail for a larger image.
Notice these screenshots are taken on a variety of platforms.

Java Tools

ControllerGUI	VisionGUI
Walk Remote Control	Head Pointer
Aibo3D	EStopGUI
Watchable Memory

MATLAB Tools

NOTE: The MATLAB tools are currently broken in version 1.5.  There is a work around: copy all of the *Listener.java files from tools/mon/org/tekkotsu/mon/ to tools/mon (with the MATLAB .m files)

Launcher	MATLAB Vision Monitor
Walk Remote Control	Joint History Graph