Most researchers developing artificial intelligences are working towards a goal of making robots that can one day adapt to any situation. But researchers at MIT have instead created an AI that can learn to drift an RC car, which means that one day the terminators will drive even better than Ken Block can.
The machine learning software is first taught how to control a standard RC car—turning the wheel to the right makes the car go right, etc.—but then has to rely on simulations to determine what corrections need to be made to acceleration and steering to account for a decreased lack of friction between the floor and tires.
Small plastic balls attached to the RC car and cameras in the lab allow the learning software to keep track of the position and motion of the vehicle while it applies what it learned in the simulation to the real world. Not surprisingly, the RC car performed slightly differently outside of the perfectly-controlled simulation, but the software was able to easily adapt its controls to ensure the vehicle wouldn’t spin out.
The researchers were even able to change the friction between the RC car’s tires and the floor using salt, but that didn’t trip up the software AI. It was also able to easily maintain control of the car when it occasionally collided with another RC vehicle. Eventually this approach to autonomous learning could be incorporated into self-driving vehicles. So when the surface of the road becomes more slippery due to weather, the vehicle will be able to quickly adapt to any scenario while maintaining complete control.
[MIT AeroAstro via IEEE Spectrum]