I studied Tilden's design of "neural network" when I was a young student. Even in early 2000's, his designs were popular, and he published a book. This particular robot was designed in the 90's, and I believe he had a more complex, spider-looking robot later.
What I found fascinating was that even with a few "analog neurons" he could engineer a complex behavior that was driving the walking gait, obstacle avoidance, and in case of solar-powered bots, sunlight seeking (thus sustaining themselves).
In this age of trillion transistors and massive GPU-powered neural networks, his designs may seem almost trivial. However, I always appreciated the clever thought that went into such a simple but effective design.
It's surprising how few rules are required for very complex emergent behavior. In the early 2000's I wrote a simulation of autonomous "agents" following a Brooks subsumption architecture. The individual bots had, IIRC, about 4 rules and the goal was to get them to collectively push a load off the screen. There were 90 agents and after starting the system, they would typically accomplish that goal in less than a minute.
I wish I'd had the funding to actually build the physical agents since that would have been far more eye opening -- would be a lot cheaper to do it these days with the availability of cheap arduino clones, etc. -- but it still made for a cool class project.
Indeed, the way emergent behaviors arise from simple rules is interesting, but to the best of my knowledge, analytically intractable for all but most simple cases. The converse way (from desired emergent behaviors to the engineered simple rules) is also intractable, perhaps even more challenging than the other way.
If you know existing literature on these topics, I would appreciate some pointers.
I haven't been involved in this stuff for quite some time: my career has been more about specific types of automation rather than general robotics, so unfortunately I don't.
Nv Neurons are easy to breadboard if you buy the specified chip, that’sa good starting point for experimenting with this design since you can hook up some LEDs to see the alternating activation. Hooking up motors and sensors will be more complicated since you need some way of driving the motors.
20 comments
[ 4.6 ms ] story [ 20.9 ms ] threadFrontpage 6 will hopefully blast those silly react or angular pages from the web.
Extremely cool looking robot and awesome site out of time. Most links even work.
Mark Tilden explaining Walkman (VBug1.5) at the 1995 BEAM Robot Games part 2 https://www.youtube.com/watch?v=7LjvzjCnv-c
More: https://www.youtube.com/watch?v=7ncDPoa_n-8&list=PLGrDwNpCsq...
What I found fascinating was that even with a few "analog neurons" he could engineer a complex behavior that was driving the walking gait, obstacle avoidance, and in case of solar-powered bots, sunlight seeking (thus sustaining themselves).
In this age of trillion transistors and massive GPU-powered neural networks, his designs may seem almost trivial. However, I always appreciated the clever thought that went into such a simple but effective design.
Is anyone but Solarbotics working on this anymore? ( they also seem to just sell the kits of robots made in 20-years ago )
http://faq.solarbotics.net/nvnet.html
I wish I'd had the funding to actually build the physical agents since that would have been far more eye opening -- would be a lot cheaper to do it these days with the availability of cheap arduino clones, etc. -- but it still made for a cool class project.
If you know existing literature on these topics, I would appreciate some pointers.