Self-Driving Cars: A history in the making since 1980’s

Do you remember the first time you heard of a self-driving car? Well, we heard it back in 2009 when 谷歌 announced its self-driving project. Cut to today, there are several companies and startups that are working in this field. Right from the camera to the full technology stack, there is no stone left unturned. Rather, there is no street left unmapped!

That is exactly what we intend to explore in this piece. How did autonomous vehicle technology reach where it is now? How has technology evolved through the years? How are startups driving innovation?

Tracing our steps

Even though self-driving cars seem like a recent phenomenon, researchers and engineers have been working on this technology for well over three decades. Autonomous driving technology finds its roots in the 1980’s when the first autonomous car was built at Carnegie Mellon University. ALVINN1, the retrofitted Army ambulance was driving around?the campus without any human intervention.

The early 2000s saw the first applications of Generation 1.0 autonomous systems. This technology relies on pre-programmed rules and sensors to navigate their environment. Imagine following a recipe where you go step by step, but the result might not be the same every time.

One of the most well-known application of Gen 1.0 autonomous driving technology are the DARPA?Grand Challenges of early 2000s. It is a prize competition for autonomous vehicles funded by the Defense Advanced Research Projects Agency, a research organization in US Department of Defense.

Stanford University's racing team won the 2005 competition. Interestingly, they were the flagbearers of Google’s first self-driving effort that began in 2009, yet another well-known application of Gen 1.0 technology.

However, the picture today is very different. There are autonomous vehicles on the streets and systems in factories. The sector is attracting mountains of capital. Coupled with increased interest in Generative AI, people hope of technological breakthroughs.

We are now witnessing the development of Generation 2.0 autonomous systems.

But what are the key differences between Gen 1.0 and Gen 2.0 technologies?

The building blocks

Evolution in autonomous systems is determined by design and management of the 3 building blocks of a system:

Generation 1.0 autonomous systems (2000-2015)

Gen 1.0 systems are developed on Convolutional Neural Networks; a neural network architecture2 that is primarily designed for computer vision tasks like image/object recognition and classification. At the time, CNNs were at the forefront of the technology, a deep learning3 “tip of the spear”.

In Gen 1.0 models, the 3 building blocks, sensing, planning and acting are developed independent of each other (see figure 2). Each block is powered by its own deep-learning model, and then integrated with the other blocks as required. This makes upgrading and scaling the system more difficult. Imagine an office where people work in silos and communicate via emails.

The sensors and chips used in these models are less mature, slow and expensive. To build autonomous cars with this technology, one has to map all areas of operation in HD. The Google self-driving car program, Waymo is based on Gen 1.0 system. If it were to come to a city in India, it would first have to do an HD mapping of the whole city before letting the car on the road. A few others that used Gen 1.0 systems were Uber ATG, Cruise and Vicarious Robotics.

Generation 2.0 autonomous systems (2016 onwards)

Generation 2.0 autonomous systems are built on transformer models, an evolution of CNNs. This approach is more flexible and faster than CNN. The drawback however is that training a transformer model requires more data and power (compute). Unlike a Gen 1.0 model where the 3 blocks were developed independent of each other, Gen 2.0 attempts to develop a common underlying deep learning model (See figure 3). Think of a co-working space where everyone interacts with each other, has their own opinions but is working towards a common goal.

The Gen 2.0 model is more difficult to troubleshoot because it is one large model and finding the cause of a glitch takes more time. It uses a more end-to-end approach. It is trained using reinforcement and imitation learning.

As for its sensors, they are cheaper and more capable, as are the chips. Unlike Gen 1.0 systems, HD mapping is not a requirement which means the model needs to be fed more data and requires more compute to work well. Some systems developed with Gen 2.0 are Tesla FSD, Covariant Robotics and Wayve .

You can also watch this lecture by Oleg Sinavski from Wayve who explains the evolution of self-driving cars.

Next generation autonomous cars

Now that we know how technology has evolved, what impact will it have on the next generation of cars? What will be the real-life implications? Read below…

• It will be easier to interact with these systems through voice. This means that soon, one can interact with an autonomous car through a text/voice interface (Read about LINGO-1 ). The system will also be able to reason and justify its actions.
• Autonomous technology has the potential to scale faster under Gen 2.0 versus Gen 1.0. One key reason for that is the car might not need exhaustive HD maps for the entire world. Like humans, once it learns to drive, it will be able to drive everywhere. Watch this video .

So how will a car look like in 15 years? Amazon’s Zoox provides a glimpse into it, a car that is reimagined and built from the ground up.

However, these amazing implications are not without concerns of their own. There is still a long way to go till autonomous cars find their way on roads commercially. Some of them are listed below:

• Transformer models require a humongous amount of high-quality data to prevent any bad correlations. Will we run out of data?
• All we have seen till now are demos which are slick and impressive. But the gap between demos and city roads is wide, therefore there is a higher chance of failure.
• These autonomous systems need to demonstrate adequate safety mechanisms before they can be launched in human environments.
• Commercial viability is under question till these models hit the real world. Till date, Tesla FSD or ‘driver assist’ is the only large-scale commercial application of Gen 2.0 and that too, has many critics.

From where we sit, the autonomous vehicle market holds immense potential. The transition from Gen 1.0 to Gen 2.0 technology signifies a more efficient path to commercialization, as it moves beyond the limitations of HD mapping. However, challenges like data security, real-world robustness, and safety remain.

Startups that can address these issues and drive responsible development are pivotal. Their ability to bridge these gaps and bring innovative solutions will attract investment and shape the future of autonomous driving.

Investing in this space offers substantial opportunities, provided one navigates the technological and practical challenges inherent in moving from impressive demos to real-world applications.

Index

1. ALVINN: It stands for Autonomous Land Vehicle In a Neural Network.
2. Neural network architecture: It is the overall structure and design of a neural network. A neural network is a blueprint for how the network is built and how information flows through it. It is a computer system inspired by the human brain, designed to learn from data.
3. Deep learning: A type of AI inspired by the structure and function of the human brain. It uses artificial neural networks. Just like our brains learn from experience, deep learning models learn from vast amounts of data.

Recommended video

Watch Huawei’s self-driving car drive on Shenzhen roads.

?