The carmaker is using text-to-image AI to produce vehicle designs that conform to engineering and manufacturing constraints; combining it with digital twin simulations means fewer iterations and faster time-to-market.
It’s no secret the automotive sector is racing to find ways of tapping into the potential of generative artificial intelligence (GenAI) to design and build the next generation of vehicles. This technology has much promise, from redefining manufacturing processes to helping carmakers design smarter, safer and more efficient vehicles.
Much of the time, proprietary automotive innovation is kept under lock-and-key as a critical competitive advantage, but recently Toyota has shared the development of a new tool that is enabling designers and engineers to collaborate more efficiently and easily.
GenAI is a type of artificial intelligence that doesn’t just focus on processing data: it uses advanced machine learning techniques—particularly deep learning—to generate new content. The technology has the potential to enable carmakers to optimize vehicle designs and structures, producing lighter, more aerodynamic and more fuel-efficient vehicles. However, the GenAI is still in its infancy and has encountered challenges evaluating complex variables such as manufacturing limitations and detailed safety regulations.
“Generative AI tools are often used as inspiration for designers, but cannot handle the complex engineering and safety considerations that go into actual car design,” said Avinash Balachandran, director of the Human Interactive Driving (HID) Division at the Toyota Research Institute (TRI). TRI is a division that focuses on incorporating next-generation technologies into the carmaker’s manufacturing processes.
TRI recently shared a GenAI process that could overcome those limitations to assist vehicle designers. These designers can already use publicly available text-to-image generative AI tools as an early step in their creative process—but TRI’s new technique combines early design sketches and engineering constraints into the process. Reconciling design ideas with engineering constraints early in the process results in fewer iterations to reach the final design.
For example, Toyota’s designers introduce constraints such as drag, which impacts fuel efficiency, into the generative AI process. Subsequent iterations would optimize drag within the parameters defined by the designer.
“This technique combines Toyota’s traditional engineering strengths with the state-of-the-art capabilities of modern generative AI,” said Balachandran. “It was motivated by the advancements in text-to-image generative AI tools, where you could type in a prompt, and it generates an image adhering to the stylistic guidance of that prompt. The inspiration for this technique and these tools was not to just spur creativity, but also to shorten that iteration loop between engineering and design,” he says.
Balachandran’s team had to tackle the difficult task of reconciling a sleek and elegant design with the realities of engineering performance and safety requirements. Designers and engineers often have very different backgrounds and ways of thinking about how a vehicle looks and performs—requiring a significant amount of back-and-forth between them to achieve a feasible solution, which can slow down the design process.
“To overcome these limitations, we built an AI model that can incorporate precise engineering constraints—like minimizing aerodynamic drag—to maximize the performance of these potential cars,” said Balachandran. “This will cut down on the number of iterations considerably and allow designers and engineers to work more closely and quickly.”
Adding those engineering constraints to the generative AI model allows the user to set limitations on the AI’s generative designs, requiring it to apply those constraints to the design. As a result, the generated design will account for factors that improve performance, safety, and reliability while satisfying the designers specific needs.
By the time the vehicle design goes to the engineering team, some of the job has already been done. “Reducing these iterations allows for faster vehicle design processes as well as improved efficiency for the design and engineering teams,” said Balachandran.
The technique has the potential to significantly accelerate electric vehicle (EV) design, in particular. “If you have superior aerodynamics, you can improve the range of that vehicle without increasing the size of the battery,” said Balachandran. “This is powerful, as large batteries are not only expensive to make but also use the limited resources that we have to build them. By focusing on drag first, we hope that we can make a big difference in the design of EVs…At the end of the day, we hope that these tools can offer value for any vehicle design though we targeted drag first as it has an outsized impact on EV designs.”
The technology can factor in any measure that is inferable from the image itself—including drag. In fact, drag is inferable because shapes have particular drag coefficients that the AI can measure. Other factors that impact ride handling, such as the wheelbase and ride height, can also be optimized by the AI.
It strikes a balance between amplifying the designers’ capabilities and the engineers’ constraints. “We spent a lot of time working with designers to understand their pain points so that we can develop techniques that added value to them,” said Balachandran. His team focused on ways the AI could assist designers by helping them focus on the parts of their job where they could apply their creativity to the fullest. They discovered that multiple iterations between the designers and engineers posed a significant challenge because it took them away from the creative process where they could add the most value—and they enjoyed the most.
Toyota’s generative AI tool also creates digital prototypes of vehicles, which are put through simulated real-world test, enabling engineers to identify potential flaws early in the development process, avoiding potentially costly flaws during production.
For example, a designer can request that the tool design a vehicle based on an initial prototype sketch with qualitative parameters such as “sleek” or “like an SUV.” The tool would interpret the request and create a few designs as requested—while still optimizing quantitative performance metrics such as aerodynamic drag.
“We’re leveraging generative AI tools that are trained on thousands of other images of vehicles,” said Balachandran. “Part of the power of these tools is that they can use the knowledge gleaned from this corpus of data to help a designer explore this subjective space and push themselves creatively.”
The tool is currently being used for vehicle handling characteristics such as drag, ride height, chassis position and structural integrity. Balachandran’s team is working with its partners across Toyota’s network to enable designers to incorporate the technique into their own workflows.
“The hope is that, by using this tool, they can expand the power of design ideas while at the same time drastically improving the speed of design development,” said Balachandran. “Generative AI is a powerful new tool. We are exploring, across our many research areas, how to leverage it responsiblyso it can amplify our people.”