A Spiking Neural Network Saved This Robot From Falling

Despite the name, most artificial neural networks (ANNs) actually have very little in common with their biological counterparts. While they are loosely inspired by the interconnected neurons found in the brain, modern neural networks are primarily mathematical models designed to recognize patterns in data. Biological neurons, on the other hand, communicate through complex electrochemical processes and adapt continuously.

A somewhat lesser-known type of ANN called a spiking neural network attempts to more closely model biological networks. Rather than processing information as continuous numerical values, spiking neural networks communicate using discrete electrical pulses, or "spikes," similar to the way neurons exchange signals in living organisms. Information is encoded not only by whether a neuron fires, but also by the timing of those spikes, allowing these networks to capture temporal relationships more naturally.

Becky Stern recently picked up a Backyard Brains SpikerBot, which is controlled by a spiking neural network. She made a few modifications to the bot so that it would be able to use its artificial brain to detect the edges of a tabletop to keep itself from falling off. Fortunately for us, Stern also documented the build so that we can all learn a bit more about this interesting neural network architecture.

SpikerBot was developed as an educational platform for teaching neuroscience concepts through hands-on experimentation. Instead of writing traditional code to define robot behavior, users create and modify simulated neural circuits that process sensor inputs and generate actions. The robot's behavior is based on a computational spiking neural network model derived from work by neuroscientist Eugene Izhikevich, allowing it to respond to its environment in a way that feels less deterministic than conventional robots.

Inside the robot is an assortment of sensors and actuators designed to mimic some of the capabilities found in living organisms. The hardware includes an ESP32-S3 microcontroller, an OV2640 camera module, a VL53L4CD laser distance sensor, a microphone, an accelerometer, RGB LEDs, and continuous-rotation servo motors. The camera can identify colors and objects, while the LEDs provide a visual indication of neural activity within the simulated brain.

Although the robot already includes a forward-facing distance sensor for obstacle avoidance, Stern wanted it to recognize when it was approaching the edge of a tabletop. During a discussion with hardware designer Alex Hatch, a simple solution emerged: redirect the laser sensor toward the floor using a mirror mounted at a 45-degree angle. To accomplish this, Stern used snap-fit accessory templates provided for the open source robot and designed a plow-shaped attachment in Tinkercad.

After 3D printing the attachment, Stern cut a small mirror to size and mounted it to the angled face of the plow. The mirror redirects the laser beam downward, allowing the distance sensor to continuously monitor the floor in front of the robot rather than obstacles directly ahead.

Stern also needed to adjust the robot's neural circuit configuration. Starting with the included Explorer brain, she inverted the distance threshold used for obstacle detection. Normally, the robot reacts when an object is close. With the sensor now pointed at the floor, the floor itself becomes the expected close object. If the sensor suddenly reports a much greater distance, the robot interprets that as the edge of the table and triggers its response.

For more details on this project, be sure to check out the video below.