As of June 2024, almost 4,000 automobile collisions involving autonomously driven vehicles had been reported in the United States (3,979 to be exact), leaving some consumers in doubt as to the reliability and safety of autonomous driving systems (ADS) and advanced driver assistance systems (ADAS).
For every 1 million miles driven, 9.1 self-driving car crashes occur. Some incidents have made headlines, such as when a Pony.ai driverless vehicle collided with a sign in 2021, a Cruise automated taxi stopped in the middle of a left turn in 2022, and another Cruise car rear-ended a bus in 2023.
Challenges with AI-Driven Autonomous Vehicles
As the tech for AI-guided driverless vehicles continues to develop, it faces several challenges. One is simply public perception, as manufacturers and sellers of these vehicles work to build trust with consumers and change the way people perceive vehicles that have AI “behind the wheel.”
A second challenge is the need to make the AI engines driving these cars more adept at handling edge cases. Edge cases are unpredictable scenarios like unusual road conditions or unexpected behavior from pedestrians. A human driver who is alert, sober, and responsible can react quickly to a pedestrian jumping unexpectedly in front of their vehicle or to unanticipated potholes, or to the presence of a bit of items in the lane that have fallen off the back of a truck, for example.
First-generation ADAS systems have had more trouble with unpredictable scenarios outside the data set the AI was trained on. End-to-end (E2E) models are optimized for the average case, so edge cases are excluded from their learning.
“When you look at autonomous driving,” Sophia Eichler of autonomous driving tech developer Autobrains explains, “you really see that AI needs to solve real-world challenges that have a lot of layers of complexity.”
Third, driverless vehicle developers face significant challenges with cost and scalability. The development cost has to be lowered – without sacrificing safety or stepping into ethical gray areas – so that the cost to the consumer can be reduced if AI autonomous vehicles are to become viable for widespread adoption and use.
The problem is that retraining and running more advanced E2E models consumes more resources with each iteration, not less; sometimes, the increase in resources expended is exponential rather than incremental.
Why Liquid AI Might Offer a Solution
That is why the most recent AI solutions are moving beyond the traditional rule-based ADAS. Liquid AI is a technological approach that divides the complex work of autonomous driving into small segments, each with its own optimized AI model that has been skilled to address a specific driving scenario. These smaller models are modular and designed to handle edge cases, and their training requires logarithmic rather than exponential growth in resources.
Autobrains, a 2019 spin-off from Portica, is developing liquid AI of this kind. It has raised $140 million from investors such as BMW and Toyota for “real-world, self-learning AI” for autonomous vehicles. Autobrains’ system is a network of small, specialized E2E models that uses binary self-learning signatures to collect data about the car’s external environment without storing personal data or imagery (thus alleviating privacy concerns).
The car’s AI imports the data into a virtual environment, where the AI can study and adapt to the environment in real time. This new model can engage in unsupervised learning and pattern recognition, potentially providing a more responsive and safe autonomous driving experience.
The Future of Driverless Cars
Because of the challenges of developing AI that can adapt to edge cases, building consumer trust in the technology, and balancing cost and scalability, it remains uncertain precisely when autonomous vehicles will become more mainstream.
Yet the newest tech in this space has promising potential. Suppose the next generation of driverless cars can anticipate and react to unexpected scenarios. In that case, it isn’t farfetched to suppose that the number of ADAS-related automobile collisions may drop, and consumer confidence in driverless cars may rise.
