I
It had very poor IR range sensors, bump sensors, and wheel rotation sensors, all kludged onto a RC "Scorcher 6x6" chassis. With fellow student Eric Bax, I taught it to follow walls (sort of) by driving it around by remote-control while collecting sensor data and recording my actions, then training a neural network to reproduce my average sensor -> action function, such as it was. My ultimate goal was to apply techniques from genetic algorithms and Kolmogorov complexity in combination with on-line, 24 hours a day operation.... didn't happen.
Send comments to
winfree@hope.caltech.edu