Autonomous Vehicles

At some point, every self-driving car needs the skills of a F1 driver.

Autonomous systems such as self-driving cars require several stages of information gathering and processing to replace a human driver. First, the system must work out where it is and what is around it (e.g., the middle lane of a highway with a truck on my left). Next it must decide what to do (e.g., stay in this lane for a bit). Finally it must decide how to do it (e.g., maintain current steering angle for a further 1.5s then ramp it down to 0 for a further 0.3s . . . etc.).

Our system will address the last of these three steps.

In the example above, this amounts to maintaining speed and steering to stay in the middle of the lane; this is easy, cars have been able to do this for years. What is not easy is being able to come up with an optimal set of controls for the vehicle (controls being primarily throttle, brake, and steering) for every possible desired manoeuvre under any possible set of conditions.

In other words, staying in a lane on a sunny day is easy, swerving to avoid a child on a snowy road at night is hard.

Hard examples of this problem may seem few and far between but if a car is aiming to achieve Level 5, or even Level 4, autonomy then it must be constantly planning emergency manoeuvres to put into effect immediately should the need arise.

The reason that swerving to avoid a child on a snowy road is much harder than staying in a lane on a sunny day is because of non-linearity. Tyres can only provide so much grip before they start sliding, whether due to braking, turning, accelerating, or a combination. When the tyres are a long way from these limits they behave linearly, that is to say a bit more turning will produce a bit more force in proportion to the amount of force that the last bit of turning produced. When the tyres are at or near the limit of grip this is no longer the case, an extra bit of turning could produce a very small amount of extra force, or it could produce less force than before.

This is why driving a car on the limit of grip is very hard. The only people who do it professionally are racing car drivers and they are far more skilled than you or I precisely because it is such a hard thing to do.

In order to guarantee that they can handle any situation, Level 5 autonomous cars will have to be constantly planning manoeuvres which require handling at the limit of grip. This is hard, requiring a robust method of solving a very non-linear optimal control problem.

Canopy is the industry leader in solving this problem in motorsport, and we’re now applying our knowledge and proprietary collocation solver to the autonomous vehicle domain.