How not to ride a bicycle
Left to its own devices, a bicycle won’t get far.
What happens to moving bicycle without someone to steer it? It falls over, of course, but the interesting questions are how far it gets and the path it takes on the way.
The image above shows the result of 800 unsteered runs of a bicycle simulator created in 2004 by computer scientist Matthew Cook, then at the California Institute of Technology. In each run, the bicycle is being pushed from left to right, and the path taken by the front wheel until the bicycle falls over is shown. The bicycle is moving too slowly to keep itself stable, which accounts for the oscillations from side to side.
It was created in the course of a project that aimed to teach a computer to ride a bicycle. The problem is more difficult than you might think: we may be able to ride bicycles ourselves, but, as Cook notes, “we do not have great insight into how we ride a bicycle”.
In the simulation, at least, Cook found that the simplest method of effectively steering the bicycle was a neural network composed of only two neurons: one to calculate the required lean of the bike to execute a given adjustment in direction, and the other to translate that change into an amount of torque to apply to the handlebars.