Hardwired for Intelligence

If you want to add some intelligence to a robot these days, the best way is to load it up with sensors that feed into an artificial intelligence (AI) model running on a small computing platform, like an NVIDIA Jetson or a Raspberry Pi. Or if you are on a budget (of either an energy or financial sort), an optimized AI algorithm running on a microcontroller will get the job done as well. The scale of things changes, but the solutions are essentially the same.

We have not always had the luxury of being able to run complex AI algorithms on low-power, portable platforms, however. YouTuber ALTco recently dug into how these problems were sometimes tackled a few decades ago, before all of our current modern conveniences were available. Using BEAM robotics as an example, ALTco discussed how very simple analog circuits — called neurons — can be arranged to bring about complex emergent behaviors.

These individual neurons only do very simple things, like oscillate between high and low voltage levels. They may also have some sort of time delay between pulses of signals. Rather than being controlled by a microprocessor, each unit is equipped with basic components like resistors and capacitors that do the job. Photoresistors may also be used to modify the duty cycle according to the level of ambient light.

In any case, these neurons are incredibly simple, so you cannot do much with them. Unless you stack them, that is. By feeding the inputs of one simple unit into the next, and building up a multilayered circuit, much more complex, emergent behaviors can arise. And if carefully designed, these behaviors can look an awful lot like intelligence.

Perhaps that should not be entirely surprising. After all, what has just been described is essentially a simple feedforward artificial neural network. Of course in this case, the neurons do not learn their “weights” from example data, but are instead tuned by choosing specific values for the passive components. And you do not want to try to scale this approach up. When moving beyond a toy problem, the complexity, size, and power consumption will quickly get out of hand.

But that is not what this project is about. ALTco is not trying to replace modern neural networks, but rather, is showing an interesting way to squeeze intelligent behavior out of a handful of resistors and capacitors. And when you look at it from that angle, what ALTco accomplished with the robot he built is pretty amazing.

ALTco’s tiny, hardwired robot uses a photoresistor and a series of other neurons to control its motors. While each neuron does almost nothing, together they give the robot the ability to navigate around an apartment. And it is not just a simple light-following robot. Sometimes it moves toward light, but not always — depending on the internal states of the neurons. The robot has also been shown to have a knack for getting itself out of tough spots right when it looks like it is about to get stuck.

Be sure to check out the video for more details on the operation of different types of hardwired neurons, and also to see how the robot was built.