Francesco De Comitéhave described his interesting robot platform based on PIC18F2620 microcontroller with 64K Flash memory and 1K of EEPROM. Such memory allows to implement some sophisticated learning schemes like reinforcement learning, neural networks implementation, population based incremental learning.

The robot itself is built on DVD box skeleton. Motion is created by using two DC motors controlled via TC4427A MOSFET drivers. Robot uses several sensors to read environmental data. These include Light sensors and proximity sensors made of IR LEDs. Robot motion can be based on reflex behavior when it moves towards light or can learn to move. This is more interesting part because it uses simple Q-learning algorithm where robot can learn to move. Algorithm simply gives a reward each time when robot’s received light intensity increases. Also robot gets rewards when robot chooses to move straight instead of spinning. In other words it is simple neural network with learning capabilities. So robot behavior is a little bit unpredictable. Program source is here.