Some of us might have experienced pieces of autonomous driving when stepping in new(er) cars, ranging from adaptive cruise control to partly autonomous driving with Tesla models. In this article I would like to go through a small history of the Tesla Autopilot and how it’s approach has been shifted and drastically accelerated using the integration of AI models.
(Greg, 2021) As of 2015 the first version of autopilot was released overnight to existing Tesla users with the right specifications. This version was simply able to drive like most adaptive cruise controls are currently able to do namely: stay within the right lane and follow other traffic. As of 2016-2019 we saw an improvement of this autopilot (enhanced autopilot), introducing automatic lane shifts, parking and more which relied heavy on thousands of lines of static code. As of 2020 Tesla has introduced Full-Self-Driving (FSD) beta’s (Greg, 2021).
As of from 2020 we have seen giant leaps forwards in regards to the capabilities of FSD, this is mainly due to the fact that Tesla has been able to convert its collected data and code based on the earlier models from 300K+ lines of code into an AI neural network (Ali, 2024). Which means that the cars no longer stop for a red light because there is a line of code stating you need to stop for red, but it determines it based on the thousands of hours of video data it has been trained on from other tesla drivers. This results in some new emergent behaviours, as for example, a youtuber was driving a tesla, but due to roadwork it decided to do a U-turn and find a different route, which according to Ashok Elluswamy (head of AI/Autopilot at Tesla) was behaviour the FSD AI had learned by itself.
Moving away from static lines of code to an AI will/is showing to make massive leaps towards a future without necessary interventions of humans when driving. Personally I like driving, but not having to will result in much time more effectively spend.
Currently there are still lots of legal and technological constraints before full autonomous self-driving will be rolled out to the public, but the future potentially looks driver-less.
Ali, I. (2024, September 9). Tesla (TSLA) reveals FSD 12.5 roadmap ahead of the Robotaxi event, v13 in October, aims FSD for China and Europe in 2025. Tesla Oracle. https://www.teslaoracle.com/2024/09/08/tesla-lays-down-critical-fsd-milestones-ahead-of-the-robotaxi-event-aims-fsd-for-china-and-europe-in-2025/
Greg . (2021, October 9). A timeline of Tesla Autopilot: From inception to now. That Tesla Channel. https://www.thatteslachannel.com/a-timeline-of-tesla-autopilot-from-inception-to-now/
Great explanation and overview of the different stages of autonomous driving! The shift from static lines of code to an AI-driven neural network is fascinating. It is interesting to see how Tesla’s Full-Self-Driving has progressed. Not just following pre-programmed instructions but by learning and adapting through thousands of hours of real-world driving data. The example of the AI learning to perform a U-turn on its own showcases just how far this technology has come. While there are still legal and technological hurdles to overcome before fully autonomous driving becomes mainstream. The progress being made in AI neural networks looks to point towards a driverless future. Personally, it is exciting to think about how more efficient our time could become when we are no longer required to manually drive. While I enjoy driving, I am looking forward to see what the next stage brings!
Interesting topic! The explanation on the progress that Tesla has made in autonomous driving is well written. The shift from relying on static code to AI neural networks really shows how much potential there is in AI-driven systems, and how rapidly things are advancing. Whilst it is amazing that an AI is able to learn how to perform a U-turn and I agree with Jurre that it provides a bright future there are indeed some hurdles that have to be tackled. Who is responsible if an accident happens because of the AI and even while the AI can learn from data, it’s not perfect. What happens if it encounters a situation that’s drastically different from its training data? Nonetheless, autonomous driving has a bright future and im curious about the developments Tesla is going to make.
I really enjoyed reading your blogpost, particularly how you elaborated on the history and current advancements of Tesla’s Autopilot. I am also fascinated by its capabilities and how the Full-Self-Driving (FDS) capabilities have shifted from static lines of code to AI-powered decision-making. Looking back, the first time in an autonomously driving car was truly an experience for me! Your examples of emergent behaviours, like the U-turn, demonstrates the potential of AI learning in real-world scenarios. However, even with this progress, I do believe that there is still a long road ahead for these systems to becoming entirely foolproof, especially in an unpredictable environment in which they operate in. I think by now, most people are aware that there have been reoccurring scenarios in which FDS caused crashes instead of preventing them due to failures within the system. Looking at the bigger picture, FDS still prevents more crashes than it leads to, so it’s simply not fully developed yet. Therefore, I think it’s crucial for companies like Tesla and the industry as a whole, to find just the right balance between embracing AI innovation, but ensuring clear safety measures. As you expressed, we must express these constraints before we can fully hand over the wheel to machines and look into a driverless future.
The article makes a meaningful and great discussion about Tesla’s exploration in the field of autonomous driving.
Tesla has accumulated a lot of resources in the field of autonomous driving neural networks, but I always believe that for autonomous vehicles, the way to obtain road information should not be just “pure vision”. The solution that uses millimeter wave radar and vision will be closer to reliability.