Mileage Data in AV Development
#data #AVs #commentary
Autonomous vehicles are more than just cutting-edge sensors and advanced algorithms. At their core, they rely on data — specifically mileage data. The sheer volume and diversity of this data are crucial in training, testing, and validating the systems so they can confidently and safely be used commercially. For sophisticated systems like AVs, comprehensive and diverse testing scenarios are crucial to uncover and address potential issues.
There are three main categories of mileage data — simulation, supervised, and driverless. Understanding these data is important for appreciating the current state of the AV industry and setting the stage for a deeper analysis of the challenges and opportunities ahead. In the next blog post, we will explore what I believe is the key for the industry moving forward.
1. Simulation Mileage
Simulation mileage is a cornerstone of AV development. Companies use sophisticated virtual environments to test and train their vehicles across countless scenarios. This synthetic data allows for rapid iteration and testing in a controlled setting, making it possible to expose AVs to scales of mileage that are challenging to accomplish in real life, as well as rare or dangerous conditions that would be difficult or unsafe to replicate in real life. AV companies may have 1000x more simulated miles than driverless miles.
The ability to generate vast amounts of data quickly and at low cost makes simulation an attractive tool for AV companies. It’s especially useful in the early stages of development or a new product, platform, or release, where the focus is on refining algorithms and improving the vehicle's decision-making processes. However, the quality of these simulations and their ability to accurately reflect real-world conditions are crucial factors that can significantly impact the effectiveness of this mileage.
2. Supervised Mileage
Supervised mileage involves collecting data from real-world driving, but with a human safety operator present to intervene if necessary. It is the only way to accumulate real-world mileage data for new or uncertain systems. This approach is essential for validating the systems developed in simulation, ensuring they perform well in the unpredictable and often chaotic environment of real roads.
Massive volumes of supervised driving data are the cornerstone of Tesla's technical strategy. Tesla accomplishes this by collecting data from all of their consumers who use Full Self-Driving (FSD). Ultimately, Tesla is getting high volumes of "free" supervised driving data, whereas other AV companies need to maintain an expensive fleet of AVs and operators to collect even a fraction of this data. Granted, there are also negatives to their strategy that we will save for another day.
While supervised mileage provides a crucial intermediate step, allowing companies to gather real-world data while maintaining a safety net, it comes with challenges. The high costs and logistical complexities associated with this type of testing mean it's typically used selectively and strategically.
3. Driverless Mileage
Driverless mileage represents the final step of AV testing. It involves vehicles operating without any human intervention, relying entirely on their autonomous systems. This data is the most direct indicator of an AV’s readiness for deployment, as it demonstrates the vehicle's ability to handle real-world conditions on its own.
Successful accumulation of driverless mileage is a significant milestone for any AV company. It not only validates the technology but also builds public and regulatory trust. However, driverless testing is also the riskiest and most scrutinized phase of development, with stringent regulatory requirements and high stakes in terms of public safety.
While driverless mileage is critical, it’s also resource-intensive. Companies must balance the need for extensive driverless testing with the operational costs and risks involved, making this type of mileage a key focus area as AV technology matures.
Conclusion
Each type of mileage — simulation, supervised, and driverless — plays a distinct role in AV development. Together, they form the backbone of an AV company's testing and validation strategy. However, based on the current industry standards and cash burn, the industry may need a strategic shift to really become successful. In the next post, we'll explore this hypothesis and delve into potential game-changing approaches that could reshape the AV landscape.