# clone openpilot into your home directory
cd ~
git clone --recurse-submodules https://github.com/commaai/openpilot.git

# setup ubuntu environment
openpilot/tools/ubuntu_setup.sh

# build openpilot
cd openpilot && scons -j$(nproc)

We get these questions over and over, and I figured it was worth writing a blog post to address them. Then, instead of reanswering them, we can just link to this blog post.

Why don’t you partner with car companies?

What does the word “partner” mean? Like “hook up”, it’s meant to keep things vague and ambiguous.

The simplest way to analyse this is to look at the money flow. Many partnerships are nothing more than a joint PR exercise, so I think it’s easy to say why we don’t do them (waste of time for nothing gained, but PR people do feel useful). Another type of partnership is where you give them money, and I think it’s self explanatory why we don’t do that. (hint: we don’t have money)

Which leaves the third type of partnership, where they give us money. This is Mobileye’s business model, and it’s most akin to how Microsoft sold Windows to computer makers. It involves huge amounts of speculative business development spend, which is unclear if it will yield anything. Many startups spend tons on this and never see a return. There’s no reason for us to take this risk.

But don’t you want to ship with the cars?

Sure! Think of how Linux ships with computers. Linux doesn’t have a BD arm which contacts Dell, Dell ships Linux on their servers because Linux is the best choice, and it’s what the customers want. openpilot will be similar.

This strategy is already bearing fruit. Aptera is integrating openpilot, and we have no “partnership” with them. We haven’t even really spoken to them. openpilot is trademarked by us, but as long as they acknowledge that it’s owned by comma and use it according to our safety model, they are welcome to use it. And if you fork openpilot and don’t use the trademark, you only have to abide by the MIT license.

This is how it should work. We make the best ADAS system, we make it available for free, then car companies will integrate it. No business development and no partnerships required, just a GitHub repo available for anyone in the world.

Why don’t you advertise?

Advertising makes little sense in the modern world. If you have a “linear” business, sure, you can work out your customer acquisition cost and spend a portion of it on advertising.

But ugh, who has a linear business anymore? If you have an exponential business, advertising will not really help. The spread is more similar to viruses (get it, virality?). All that matters is the R0 of the product.

The most sustainable ways to raise R0 are to:

1. make a product do more
2. make a product cost less (both money and hassle)

We focus on this instead of advertising.

Why don’t you use the cameras/screen in the car?

The built-in cameras are very hard to use. They are too high bandwidth for the CAN bus, and would require very complex install processes that differ a lot between cars. And in many cars, the camera is directly connecting to processing and there’s no way to get the raw images. They will never be usable.

The built-in screen is perhaps simpler, as you could spoof being Android Auto or Apple Carplay. The main issue here is where would you connect your phone? People do not want to give that up.

openpilot supports over 200 different makes of cars, most of which we don’t own. Adding anything car specific, even just a beeper, is very time consuming. You think these things are doable because you have one car, we have hundreds to think about.

You will never solve self driving with one camera.

Since the comma three shipped with a 360 camera, we get this a lot less. But there’s still many people who think sensors are the reason self driving doesn’t work. This is completely untrue, and a myth pushed by big LIDAR. I think people like to bikeshed about sensors because unlike software, it’s something they can understand.

Remember, humans (even deaf ones with one working eye and non 20/20 vision) are crazy good drivers. Think about how many times you have driven to work. Think about how many times you crashed. And if you ever have crashed or been crashed into, was it a failure of sensors or a failure of attention?

Do this experiment. Go on a drive. Track the times you needed to correct openpilot, even with a comma two. Watch the replays in connect. I promise over 95% of the mistakes have nothing to do with the sensors and everything to do with the software.

Your comma three with openpilot is a level 2 ADAS system. It will never be level 5. But it’s held back so much more by software and not sensors. Watching the replays, how many of the mistakes could you have corrected just watching the videos recorded by the device?