Introducing carrot model

1 minute read

Due to the big success of kale model, we have decided to make more vegetable based models.

A major problem in self driving cars is determining where you want the car to go. If you know where you want the car to go, all that’s left to do is make the car go there. We considered many solutions to this problem, some involving fancy machine learning. But we eventually settled on something simple.

In openpilot 0.8.4, which is the first openpilot release including carrot model, you can take advantage of this new feature. You just need four things:

  1. A driver to be paying attention (needed at all times!)
  2. A carrot
  3. A fishing pole
  4. A passenger standing in your sunroof holding said fishing pole
Note: wind will cause ping-pong. Keep your arms steady.

We’ve trained carrot model to chase carrots. It loves carrots! In its small conscious mind, it dreams of finally catching a carrot. (note: don’t let it do this, that is off the map, carrot works best at a distance of 1-2 meters from comma two)

But we aren’t stopping at carrot model. Our research division is investigating stick model to punish the car when it does badly, and also carrotless model, which is like carrot model, but uses a virtual carrot instead of a real carrot, easing the need for the passenger, fishing pole, and carrot. We have very high hopes for carrotless model.

openpilot gets better with every release, and with a comma two running 0.8.4, you can finally use carrots to properly motivate the model. A feature nobody asked for perhaps, but a feature you will grow with and learn to love.