Generative AI and Autonomous Trucking
Generative AI will be a game changer in transportation. What about its impact on advancing autonomous trucking?
The buzz phrase of the year, “Generative artificial intelligence (AI),” has the potential to significantly impact and improve the transportation system across many modes. From optimized traffic management, predictive maintenance, supply chain, and logistics optimizations to improved roads and urban planning, Generative AI will be a game changer in transportation. What about its impact on advancing autonomous trucking?
Why is “solving” self-driving vehicles so hard?
An autonomous vehicle is not a single system that will be solved and put into mass production in a specific year. A self-driving vehicle comprises highly tuned sensing devices, complex computer hardware, and software algorithms integrated into various vehicles to provide perception, localization, planning, and control. The development and refinement of these critical algorithms require massive amounts of data collected from different locations, times, and scenarios and extensive large language training models to safely and efficiently make real-time driving decisions in a complex and challenging environment.
Many companies, including Tesla, Waymo, Uber, Kodiak, and others, have invested millions of dollars and spent many years on the road collecting data to train algorithms. The software development cycle has been very complex, expensive, and labor-intensive. This cycle has been chiefly a hand-engineered process: drive millions of miles on the road with safety drivers; collect as much data as possible; manually triage and review what the computer perceived and decisions it made; label and refine the algorithm; and develop the next software release for testing.
Rinse and repeat many, many times.
Many innovative companies see the potential of Generative AI and are integrating new training techniques to improve this process.
In addition to using simulation to train algorithms, Generative AI might be a
game changer in other areas of autonomy.
A Canadian start-up, Waabi, funded and powered in part by NVIDIA, promises to use Generative AI to massively accelerate this process. Instead of driving on the road for many miles and collecting data, Waabi trains their system in a completely virtual environment. It uses a combination of two generative AI computer systems. One system continuously simulates and generates, in real-time, roads, cities, traffic, weather, and more, called Waabi World.
The second system, running in parallel, is a virtual simulated truck called Waabi Driver, which navigates this simulated Waabi World and is capable of human-like driving, reasoning, and decision-making. When paired together, these systems reduce the need for extensive on-road testing. This model can be run unsupervised and learn continuously and quickly in the simulated environment. Waabi promises a 10x faster development cycle at one-tenth the cost, according to NVIDIA’s website.
For Star Trek fans, think about the Holodeck creating an ever-changing and challenging virtual world from which the virtual AV truck constantly navigates and learns. With Generative AI, the simulated world can change weather patterns, introduce new and unexpected actors into environments, and replicate this in various cities and highways. It can simulate edge cases, such as emergency vehicles, unexpected construction or lane changes, multiple terrains, elevations, and weather challenges, and bad actors that unexpectedly change lanes, drive the wrong way, speed, or make wrong turns, and it can learn and evolve.
In addition to using simulation to train algorithms, Generative AI might be a game changer in other areas of autonomy. Some innovations include optimizing sensor placement and integration into the vehicles by OEMs, optimizing fuel efficiencies in driving environments and modes, testing and validating emerging technologies and sensors, integrating different types of sensors for multi-sensor fusion packages, incorporating memory modules from other similar platforms, and facilitating distributed training models.
To operationalize driverless trucks, the AV industry must ensure that the vehicles are as safe as human-driven trucks across many domains and situations. This bar is very high, and progress is rightfully deliberate and measured. However, many exciting innovations are being tested and advanced, both in closed-loop testing facilities and on the road, with safety drivers for added security. Over the next few months, we will explore some of these companies and innovations at the forefront of this exciting space.
