Generative AI could shorten software development cycles while also creating distinct, hyper-personalised customer experiences. By Stewart Burnett
Software-defined vehicles (SDVs) promise to transform the mobility landscape but realising them is no easy task. Even the launch of vehicles with partial software integration has proven rocky. For example, Volvo Cars stalled the launch of its EX90 SUV by almost a year due to problems with software development. The vehicle will finally ship in Q4 2024, albeit with fewer features than initially expected. For Volkswagen, software development has proven highly expensive. In addition to the US$13.3bn it has poured into wholly-owned software firm Cariad—only to be racked with numerous delays to vehicle launches—it has also invested US$5bn into Rivian to develop an SDV platform.
Clearly, automakers need all the help they can get, and there are many companies looking to pitch in solutions. Technology services firm Tata Elxsi—part of the wider Tata Group—believes the deep integration of generative AI tools will be crucial for any automaker looking to be competitive within the emerging SDV space. Many automakers have already experimented with generative AI in areas such as vehicle architecture prototyping. However, Tata Elxsi wants to see the technology implemented not only throughout software design at the coding and validation levels but also in a range of customer-facing applications.
Establishing the foundations
“At Tata Elxsi, we are putting in a lot of effort to help our workforce and automotive clients get ready for generative AI and SDVs,” Biswajit Biswas, Chief Data Scientist at Tata Elxsi, tells Automotive World. Regardless of application, the journey towards generative AI integration begins with laying out extensive groundwork: “This is an important transformational tool, but it also requires quite a lot of different pieces to work effectively.”
The first step is to bring in large language models (LLMs) alongside other types of AI—typically under lease from major tech companies—to conduct the fundamental task of generation. LLMs, for instance, can generate everything from written content to programming code. However, all of these require a significant amount of computing power to function. As such, Biswas recommends also implementing “Vehicle Foundation Models” (VFMs), which are specifically targeted for SDV vehicles. VFMs are core building blocks for SDV co-pilots that can aid in every process of the design lifecycle.
It is crucial to supply the LLM with an “enormous” amount of training data so that it can generate contextually relevant and useful results. This may require some work beforehand to convert data into compatible file types. Some data—such as real-world vehicle use, including the owner’s sensitive personal information—will require the careful navigation of regulations, especially if transported across borders.
Transforming vehicle design
Once these fundamentals have been set in place, a wide variety of development tools can be implemented. “With the right tools, you can fine-tune generative AI technology to make vehicle-specific resources,” remarks Biswas. These can be developed internally or outsourced from other companies. Tata Elxsi, for example, offers tools for coaching autonomous driving, as well as an SDV development and validation framework called Avenir, which automates the process of software testing and interoperability, and according to Khan, “can accelerate the SDV journey of OEMs and tier ones”.
Given the recent struggles of automakers to develop software in a timely manner, however, the most promising use of generative AI could be automatic code generation. While this could be done via a text-based prompt, Tata Elxsi believes it should be as intuitive as possible. “If we are using voice-enabled features for customer-facing applications, we can also use them to allow engineers to interact with the systems they are developing. This is far better than developing everything from scratch,” states Jihas Khan, Practice Head, Virtualization, Tata Elxsi.
He adds that this approach would save a “lot of cost and effort” for automakers and Tier 1 suppliers. At the same time, it relies on pre-existing software code and so may not be a viable tool for companies just getting started on their journey towards SDV development. The more plentiful and diverse the training data, Khan emphasises, the more useful the resulting code will be. While a human engineer is generally expected to intervene in this process by reviewing the resulting generated code, this could also be automated using LLMs. “The generative AI frameworks can be utilised for review of the source code, and then we can also use LLMs to generate new codes within the given requirements.”
Better customer experiences
Looking to the future, Biswas believes that generative AI could allow consumers to actively participate in the design of their SDV, from physical hardware and aesthetic features to the software capabilities it ships with. “People will have the option of using an app to generate designs of their car and select all the features they want. In this case, we’re also leaning towards increasing the use of digital twins.” This could be of particular interest to Gen Z, a demographic he believes wants to actively shape and experience the products it uses before they are physically produced.
Any application that is required to interact with the vehicle absolutely needs to be handled offline
Beyond vehicle design, generative AI finds a range of applications within the driving experience. The most notable example is integration with the voice assistant to enable responses to nuanced or subjective queries—for example, “What is the closest restaurant that offers food I like?” However, Biswas is also a proponent of using the voice assistant for applications beyond driver comfort. By integrating the vehicle’s user manual into the LLM’s training data, the driver could troubleshoot in real-time. “You can check with your copilot and ask about the nature of a blinking light on the dashboard or the potential causes of a weird noise coming from the engine.”
One potential issue with generative AI in customer-facing applications is the need for constant cloud connectivity. Khan notes the importance of automakers integrating 5G SIM capabilities into their vehicles at higher levels. However, even in this eventuality, there will still be connectivity dead zones. Biswas recommends that automakers try to integrate as many core generative AI features—such as the ability to relay information from the user manual—into the vehicle itself as possible. “Any application that is required to interact with the vehicle absolutely needs to be handled offline.”
Ultimately, the goal for automakers is to reduce development costs while creating avenues for recurring revenue streams. Beyond voice assistants, this could mean instantly customisable infotainment user interfaces and the provision of highly tailored media content. Biswas concludes that generative AI will become invaluable for introducing new, software-driven opportunities even as the cost of physical vehicle hardware— most notably electric vehicle batteries—increases. “These cars are going to be data centres on wheels, and that means many more monetisation options with generative AI. It is one of the most promising ways to create new sales after that initial vehicle purchase.”
