AI's Road Ahead: Autonomous Vehicle Insights

The quest for fully autonomous vehicles has seen substantial investments and progress since the idea was first introduced in the late 1980s. However, even with the advancements of artificial intelligence (AI) and machine learning (ML), solving automated driving has proved more challenging than initially envisioned as traditional ML methods struggle to comprehend the complexities of real-world interactions between vehicles and their surroundings - in other words, the models can understand the behavior of the vehicles, but then they must try to comprehend the actions of all the different objects that are in the world around them as well.

Tesla's approach, focusing on stereoscopic vision akin to human perception, emphasizes real-time identification of lane lines, stop signs, and intersections. In contrast, General Motors is pioneering a different approach centered around high-definition mapping. They create precise reference models of the world, including lane specifics, traffic signs, and other critical elements. This approach aims to reduce the vehicle's onboard perception workload by using pre-mapped data, enabling a more structured and controlled autonomous driving experience. This is particularly helpful for full self-driving vehicles and robotaxis, as well as automated driving in retail vehicles – all of which are trying to perceive the world around them.

The recent COSM conference provided a platform to discuss the intricate landscape of AI, including its implications on the future of transportation. Industry leaders contributing their expertise included Rich Pople, Executive Director of General Motors’ AI and ML Center; Dorothy Li, VP at Microsoft and former CTO of Convoy, and Mark Masongsong, Co-Founder and CEO of Urbanlogiq.

Not Surprisingly, it’s All About the Data

When it comes to the future of the autonomous vehicle industry, how do manufacturers create models that recognize every possible combination of things that happen in the world? “That's an enormous amount of data and an enormous number of scenarios that need to be planned and understood,” said General Motor’s Pople. “In our case, it means that we actually have to build specialized AI kinds of tools that exist outside of the vehicle that we can be pushing into these high-definition maps. And that’s a real challenge.”

With over 3.2 million miles of roadway in the US, the ability to create an all-inclusive reference map is incredibly difficult. GM’s approach is to marry motility data collected from vehicles with aerial photography and images to create high-definition maps via AI and ML in order to invent scenarios and models.

Li underscored the significance of generative AI and its ability to thrive on handling exceptions, which are frequent in transportation. Unlike traditional AI, which produces fixed algorithms, genAI can adapt based on real-time scenarios like traffic, vehicle condition, and weather, directly leading to streamlined operations, improved safety measures, and enhanced decision-making through advanced data analysis.

"I'm very optimistic about this next generation of generative AI being applied to solve the complicated digital physical world that we're in now,” said Li. “That said, I think lots of human judgment in the mix and figuring out where humans should be in the loop when we use ML and AI, when a complete automated solution is there and when we plug in.”

Li shared that engineers she worked with used open-source resources and traffic data to build a chatbot that, within a week, could inquire about shipment details and driver instructions. This agility illustrates how generative AI enables swift ideation, testing, and human collaboration. Li’s excitement for the symbiotic relationship between AI and humans extended to folks doing unglamorous and tedious work in the transportation industry and how genAI could increase productivity through better collaboration.

Masongsong's focus was on using data analytics for informed decision-making in urban planning. “I think what's exciting for us is there is enough data out there to be making smart decisions,” said Masongsong. “Government, they just never had the capacity to consider all the different factors that we know come into play -- machine and deep learning, opens that up.”

For instance, in San Jose, a data-driven decision support system considers 40 key performance indicators for policy decisions related to land or infrastructure. This system aids in predicting the impact of decisions on equity, economic stability, and environmental protection. Masongsong emphasized that while there's always been ample data, the challenge was lacking the computational power and models to analyze it effectively. Now, with the advent of technology, cities can optimize their decisions based on a comprehensive set of indicators, ensuring better outcomes in various aspects of city planning.

The Challenges in AI Development

When talk turned to the challenges in AI development for autonomous vehicles, the experts identified several key areas, including:

• ‍Regulatory Complexity: Pople highlighted the extensive regulations within the heavily regulated automotive industry, emphasizing the challenge in navigating compliance challenges when developing new capabilities for autonomous vehicles. The impact of errors in AI models deployed in autonomous vehicles is significantly higher, necessitating stringent safety considerations that comply with complex regulatory frameworks.
• ‍Varied Road Systems and Driving Behaviors: Understanding the context of road systems designed by traffic engineers presents a substantial challenge. Factors like regional driving behaviors, unique road customs (e.g., the "Pittsburgh left"), and varied traffic patterns across different regions add complexity to creating a universal autonomous vehicle model or a national product. Li further emphasized the diversity in weather conditions within short distances, underscoring the challenge of accommodating diverse environmental factors. ‍
• Sparse and Limited Data: Li and Pople expressed concerns about the scarcity of data, especially in trucking, hindering AI development. Li emphasized the challenge of acquiring adequate data from trucks and IoT sensors, limiting AI's training potential. Masongsong also addressed the scarcity of usable data in specific scenarios, highlighting the necessity for comprehensive and accessible government data to augment AI development.‍
• Synthetic Data Generation: Pople discussed the potential of creating synthetic data to supplement the scarcity of real-world training data. Synthetic data generation could aid where specific training data is insufficient or unavailable, in essence stitching images together from a variety of sources to create new kinds of training scenarios. However, maintaining the integrity of synthetic data and avoiding model collapse will be critically important.
• Data Governance and Compliance. Masongsong highlighted the need for mature data governance and compliance structures within governments. The discussion emphasized the importance of defining rules and regulations surrounding data access, utilization, and sharing, especially concerning government-provided data essential for AI development.

‍

How Data Usage and Data Privacy Factor In

As referenced above, autonomous vehicles represent a trove of data capturing diverse driving patterns, habits, and environmental conditions, crucial for refining AI models and enhancing system performance. The extensive data collected includes driving behaviors, route specifics, speed variations, and regional driving idiosyncrasies. Amid this, governments grapple with data privacy concerns, reflected in regulatory frameworks like GDPR in Europe and California's CFPA. Discussions often weigh the balance between public good and privacy exceptions for smart city planning or societal benefits, requiring strict controls and safeguards.

Leaders in the automotive industry underscore a profound commitment to data privacy. They prioritize customer confidentiality, employing robust measures to safeguard and anonymize collected data, including everything from GPS records to vehicle identification numbers. According to Masongsong, “most governments are basically adopting the same types of standards. I think in the transportation space it'll be a lot of noise, but I think as long as there's clear definitions of why data is being used, people will be commonsense about it.”

Beyond privacy, the human interaction and trust aspects of autonomous systems are crucial. Transparency, superior performance compared to human drivers, and incremental introductions of new capabilities are seen as pivotal in building consumer trust. Real-world experiences relay positive feedback on the comfort of autonomous driving, potentially contributing to widespread adoption once consumer confidence solidifies.

What Does the Future Hold?

Autonomous vehicles and self-driving cars are expected to be fully functioning in five to ten years. As technology continues to advance, the interplay between data usage, privacy, and user experiences in the transportation landscape remains complex, necessitating a delicate balance between progress and regulation. Further, as more and more local governments establish their own rules around what types of self-driving vehicles and data science are allowed in their cities, a national approach is increasingly necessary.

Launch is on a mission to get every organization – and government – AI-ready. How do you stack up? Take our free AI Readiness Self-Assessment to find out.

And watch this clip from COSM featuring Rich Pople from General Motor’s AI and ML Center speaking about an autonomous vehicle use case that is particularly exciting to him and how connective technology and the creation of digital twins can increase vehicle safety and efficiency.