Special Talk by Dave Kortenkamp

Worlds to Explore: Autonomy Challenges for Human Space Flight

• Date: Friday, March 3
• Time: 3:10 p.m.
• Place: 101 Roberts
• Dave's home page

Abstract: This talk will look at two distinct artificial intelligence challenges in human space flight. The first challenge can be compared to building HAL 9000 from the movie 2001: A Space Odyssey, that is, autonomous monitoring and control of long duration mission functions. The second challenge can be compared to building Lt. Cmdr. Data from Star Trek: The Next Generation, that is, autonomous control of highly dexterous robots. Both challenges will be highlighted by showing research being done at NASA Johnson Space Center. This research includes autonomous control of advanced life support system tests and control of a humanoid robot called Robonaut. The talk will include many videos of NASA's robotics research.

Face Recognition Example

• Take a look at Figure 4.10
• Input encoding: map a 120 by 128 pixel picture with intensity values ranging from 0 to 255 onto a 30 by 32 grid, using average pixel values.
• Output encoding: 4 nodes representing left, right, up or straight
• Desired classifications: Use 0.1 instead of 0 and 0.9 instead of 1
• Number of hidden units: 3
• Learning constant: 0.3
• Momentum constant: 0.3
• Some interpretation of the learned weights is possible.

Neural Network Variations

• Use different error functions.
• Use different error minimization procedures.
• Allow network to be recurrent.
• Dynamically modify network structure.