Autonomous Driving: The Mother of All AI-Problems
On Armies and Guerrillas of Autonomous Driving.
Autonomous cars were supposed to be mainstream by 2020. It’s been five years since General Motors acquired Cruise. A lot happened since the billion-dollar acquisition; Uber and Lyft closed respective autonomous technology shops for starters. Most startups merged with the big giants, and we were left waiting on the pavement for the promised robotaxi.
Let’s take a look at what it takes to build a self-driving stack. The current self-driving race was kicked off when deep learning had a breakthrough in computer vision. And progress in AI still drives that race. Training and deploying SOTA machine learning systems comes down to tight integration between collecting data and deploying optimized models for custom silicon. It is now a well-accepted fact that data and compute govern the progress in AI. Based on what we know so far about AI and self-driving, we can make an educated guess at which companies could emerge likely winners.
Not all companies are equal. We have Google, Volkswagen, Baidu, etc., with infinite resources at their disposal on the one side and startups with unique strategies and resilience as their primary resources on the other side. Hence, I classified the companies into two categories, the former as armies and the latter as guerrillas.
Winners in the Army category.
Waymo: Waymo is that one guy in the gym who still believes in doing individual body part isolation exercises per day, while the others have moved on to CrossFit, Hot Yoga, and Berry’s Bootcamp. Waymo was the first to launch commercial efforts to build a self-driving car; it probably has the most complex system in place. The experience of knowing what doesn’t work, the sensor tech build from the ground up, and expertise in machine learning makes Waymo one of the frontrunners in the race. Having Google Brain, Google Robotics, and Deepmind by its side only make it even more formidable.
Tesla: Tesla autopilot has become synonymous with full self-driving. Tesla excels at all three components when it comes to self-driving. It has the most extensive dataset sourced by an ever-growing fleet. Unlike Waymo, Tesla is present worldwide and has access to the most diverse data of any companies. Thanks to its active fleet, Tesla has a unique advantage in deploying the tech and learning from user feedback even when autopilot is not engaged. The data curation and training pipeline of Tesla is an engineering masterpiece. Tesla has a clear edge in real-world data and how quickly they can iterate the models.
Mobileye: Mobileye, an Intel company, for some reason, doesn’t get enough attention. However, Mobileye’s system is on par with Tesla. And like Tesla, it builds its hardware, has a fleet of cars sourcing data and drives autonomously in dense areas without disengagements. One of the remarkable things about Mobileye is that cameras and lidars are built to be redundant in its architecture. That makes its stack exceptionally robust to edge cases.
Winners in the Guerrilla category.
CommaAI: Comma is one of most the exciting companies in the space. Based in the US, like Tesla, it too has extensive real-world datasets sourced from an ever-growing fleet of OpenPilot. It is the only company other than Tesla that ships the tech and directly learns from every mile driven with or without OpenPilot engaged. And Comma does it with a team of 10 people by open-sourcing OpenPilot and leveraging vast and dedicated community enthusiasts and hackers.
Kopernikus Auto: Kopernikus is a self-driving company based in Germany. While the world was focused on full self-driving, Kopernikus carved out a niche by building autonomous driving solutions for production facilities and parking garages. Kopernikus uses (existing) security cameras mounted on the walls and light posts as the primary sensors. Kopernikus SW is already a part of serial production cars. When the regulations are in place, Kopernikus will be in an excellent position to mainstream its solution on public roads.
The approaches taken by the companies in the guerilla category are exciting. Be it Comma leveraging the open-source community or Kopernikus mastering infrastructure-based driving, every company has taken a unique approach and has proven to be extraordinarily innovative.
Surprise Winner!
Most problems have multiple solutions. The same goes for self-driving. A company with the most extensive and diverse data might win (Tesla?), or a company that can best integrate different sensors and systems might solve it (Apple?), or it could be a team that trains a large enough model on an equally large and diverse dataset. Taking all this into account, let’s speculate on a wildcard entry.
OpenAI + Microsoft: As the AI community witnessed this year with GPT-3, CLIP and DALL-E, features that emerged in these models resulted from the scale and were not explicitly trained for or planned. So, suppose OpenAI trains a model that somehow also “learns” the physics of the natural world. In that case, OpenAI could finetune that model with a lot less data than a typical self-driving company. Microsoft is the preferred partner to OpenAI and is a partner of choice for cloud and enterprise solutions to most automotive companies. It could quickly scale the solution from OpenAI and deliver it to OEMs.
In sum, autonomous driving is the mother of all AI problems. It has the potential to change how we perceive mobility and transportation. It doesn’t matter who solves it first. I am grateful to have an opportunity to witness this transition in a first-row seat.
Thanks, Esther, Chip, and Mike, for feedback on the earlier draft of this article.
