1859 Laboratories and Anybots conducted a joint project to evolve neural controllers for robotics. Here, an evolved neural network is controlling Anybots' Dexter humanoid robot.
We created populations of neural networks that controlled an accurate physical simulation of this pneumatic-powered robot. The task was to remain standing as long as possible. During the task, the robot was subjected to bumps of random magnitude and direction; thereby the neural networks evolved to recover from bumps. The single-layer networks in this experiment are not "clever" enough to perform more complex actions, like picking up a foot and stepping back to prevent a fall. The video shows the best single neural network, from the end of the evolutionary process, running on the real hardware.
Copyright 2009, 1859 Laboratories, Inc.
It reacts the same way.
cmpresents 4 months ago
Wow, it acts just like my robot.
cmpresents 4 months ago
If you guys are having hard time with processing power you could simulate the population on a network.
Each simulation is loaded on individual PCs, the results of the fitness are sent to a server that does the mixing and matching on the population.
Similar to video games where the coordinates, of each player is sent to one PC but the actual simulations are run on separate PCs.
AndrianGA 1 year ago
What was the neuron model that you used?
I'd love to work with you. I've been trying to get people together to do a similar thing for years.
I'm by my self, for now exploring the microcontrollers. I'd love to have more people to work on accurate physics engine and I want accurate graphics as input to the virtual cameras.
I'm planing to use PIC32 and read the neural network of a SD card (16GB). It might not be fast enough, but I'm only one guy
Now big did your neural networks grew?
AndrianGA 1 year ago
What is it's reaction to a force great enough to push it over?
snedie69er 1 year ago
haha, nice.
It moves like a ninja. ^_^
gabydewilde 2 years ago