What does a typical L4 sensor setup look like?

Recently we had an activity where the team did a crowd based search on public available information on some of the key players in the L4 market (waymo, tesla, baidu apollo.auto open source, cruise tusimple, aurora, Mobileye, einride, autox, Kodiak and embark) to figure out how they approach the L4 challenge. It was quite amazing to see how quick the team was able to figure out an amazing amount of details about the technical approaches.

One thing where we have been able to figure out a lot of information was on the sensor setups. IT’s important to understand that from my point of view there are two different approaches to L4: One approach is the “mass market vehicles” where the goal is to build hundred of thousands or even millions of L4 vehicles. The other one is the “commercial autonomous vehicle”, where the goal is to replace drivers in commercial use cases in a comparatively small fleet of vehicles for the transportation of goods or passengers.

Most of the players we had a look at fall into the second category with tesla seeming to be to only exception being closer to category one. One difference became quite obvious in the analysis: while the sensor setups might be quite cost driven in a mass market deployment, there is a massive amount of sensors installed in the “commercial L4 vehicles”. Actually, the challenge seem more to be early in the deployment of autonomous fleet accepting higher cost.

The common base of sensors are cameras, lidars and radars. The number of sensors we figured out is 6 to 29 cameras, 3 to 9 lidars with the exception of Tesla not using a lidar and 5-21 radars with the exceptions AutoX not having a radar and Tesla starting to remove the single radar around may 2021 depending on region (https://www.tesla.com/support/transitioning-tesla-vision). At least one of the lidars typically has very high resolution and range. Further sensor modalities mentioned are thermal cameras and ultrasonics. The chart shows a boxplot containing more information about the distribution of installed sensors. Of course, we neither now if all the sensors are in use nor if the setups have been changed since the publication. Still, quite a hefty amount of sensing!

Also, again with the exception of Tesla, using some kind of HD map seems to be a common standard. But maybe that’s a topic for a separate post.

Thanks to all the team contributing to the analysis: Alice Natoli, Daniel Kadletz, Linda Schubert, Andrei Ignat, Madhu Nikaash Yellapragada, Alexandru Puscasu, Matthias Komar, Niels Christmann, Julien Seitz, Ruijao Yan, Roxana Florescu, Christopher Pinke, Matthias Schreier, Stefan Luthard, Stephan Kirstein, Stefan Hegemann, Annemarie Albrecht

Some References for the sensor setups:

https://blog.waymo.com/2020/03/introducing-5th-generation-waymo-driver.html

https://blog.waymo.com/2021/12/designed-to-deliver.html

https://www.i-micronews.com/under-the-hood-a-tesla-model-3-tear-down-after-a-hardware-retrofit

https://aurora.tech/aurora-driver

https://gmauthority.com/blog/2018/10/how-the-gm-cruise-av-senses-the-world-around-it/ https://www.avl.com/documents/4053291/0/The+Future+of+Transport+-+Einride

https://www.allaboutcircuits.com/news/on-semis-image-sensors-and-silicon-photomultipliers-give-autox-a-leg-up-on-autonomous-driving/

https://www.mobileye.com/blog/mobileye-drive-self-driving-system/