Upload program from the nxt robot to dance
The neurons are addressed by IP and port number. The system uses an integrate and fire algorithm. Each neuron sums the weights and fires if it exceeds a threshold.
The accumulator is zeroed if no message arrives in a ms window or if the neuron fires. This is similar to what happens in the real neural network, but not exact. The software works with sensors and effectors provided by a simple LEGO robot. The sensors are sampled every ms.
For example, the sonar sensor on the robot is wired as the worm's nose. If anything comes within 20cm of the "nose" then UDP packets are sent to the sensory neurons in the network. The same idea is applied to the 95 motor neurons but these are mapped from the two rows of muscles on the left and right to the left and right motors on the robot. The motor signals are accumulated and applied to control the speed of each motor.
The motor neurons can be excitatory or inhibitory and positive and negative weights are used. It is claimed that the robot behaved in ways that are similar to observed C. Stimulation of the nose stopped forward motion. Touching the anterior and posterior touch sensors made the robot move forward and back accordingly.
Stimulating the food sensor made the robot move forward. The key point is that there was no programming or learning involved to create the behaviors. The connectome of the worm was mapped and implemented as a software system and the behaviors emerge. The conectome may only consist of neurons but it is self-stimulating and it is difficult to understand how it works - but it does.
Currently the connectome model is being transferred to a Raspberry Pi and a self-contained Pi robot is being constructed. The system uses an integrate and fire algorithm. Each neuron sums the weights and fires if it exceeds a threshold. The accumulator is zeroed if no message arrives in a ms window or if the neuron fires.
This is similar to what happens in the real neural network, but not exact. The software works with sensors and effectors provided by a simple LEGO robot. The sensors are sampled every ms. For example, the sonar sensor on the robot is wired as the worm's nose. If anything comes within 20cm of the "nose" then UDP packets are sent to the sensory neurons in the network. The same idea is applied to the 95 motor neurons but these are mapped from the two rows of muscles on the left and right to the left and right motors on the robot.
The motor signals are accumulated and applied to control the speed of each motor. The motor neurons can be excitatory or inhibitory and positive and negative weights are used. It is claimed that the robot behaved in ways that are similar to observed C. Stimulation of the nose stopped forward motion. Touching the anterior and posterior touch sensors made the robot move forward and back accordingly.
Stimulating the food sensor made the robot move forward. The key point is that there was no programming or learning involved to create the behaviors.
The connectome of the worm was mapped and implemented as a software system and the behaviors emerge. The conectome may only consist of neurons but it is self-stimulating and it is difficult to understand how it works - but it does.
Currently the connectome model is being transferred to a Raspberry Pi and a self-contained Pi robot is being constructed. It is suggested that it might have practical application as some sort of mobile sensor - exploring its environment and reporting back results. Given its limited range of behaviors, it seems unlikely to be of practical value, but given more neurons this might change.