How to leverage Generative AI for Automotive Software?

The rapid progress made recently in the area of GAI (Generative Artificial Intelligence) has prompted all the industry sectors to explore and evaluate how GAI can be applied for improving their business.

In the mobility sector, AI and neural networks have been the foundational elements for the use cases around automated driving like identifying objects, monitoring targets, and charting paths. Advances in Generative AI have however added a new and additional dimension where images, audio and text can be generated with advanced context processing using supervised and unsupervised learning.

Data samples can be generated using models like variational autoencoders (VAE) and generative adversarial networks (GANs). Adaptive foundation models often use self-supervised learning, seen in autoregressive transformers like GPT-3.5, for coherent output. Reinforcement learning enhances complex behaviours by fine-tuning models through a supervised reward system, thus sharpening the generated data.

Frameworks like LangChain and Huggingface can be used to build powerful GAI applications which can cater to various use cases for the mobility sector.

Today companies can even build their own custom GAI framework using various components as shown in a very simplistic high- level architecture below:

We will touch upon few potential use cases below which can be catered to using GAI.

This can help the reader better understand the potential of GAI for various Automotive software domains.

A Multi View and Weather autonomous driving dataset built using AI generated content

On-road driving poses numerous challenges to autonomous perception and planning systems, extreme weather conditions lead to poor visibility due to fail detections from surround view cameras. Synthetic datasets generated using vision-based GAI foundation models with prompts can be used to improve the adversarial robustness of the perception system.

This dataset improves the state-of-the-art detectors, classifiers and segmentation architectures by increasing their performance and generalization ability.

When the AI model is trained on such data and deployed in the real-world condition, it would perform better as it has inherently learnt the variations from the data provided.

Harnessing GAI models for Efficient 3D Design and Development of Automobile Systems and Subsystems

GAI models can be used to blend a text-conditioned diffusion model with a 3D reconstruction model. By leveraging acquired data, one can make a model with informed assumptions about new designs and specifications, utilizing text prompts as input.

The diffusion model generates remarkable 2D images through this process, subsequently reconstructed into 3D representations.

This makes it possible to fine-tune 3D models precisely, guaranteeing their authenticity and conformance to real-world proportions, which can enable the mobility sector to achieve new levels of efficiency, accuracy, and competitiveness.

The use cases shared above are only an early indicator of GAI’s potential in shaping Automotive software applications.

Companies are expected to adopt GAI within their software development as well as end-user application domains.

GAI has created new excitement and expectations in the market, and it will be very interesting to see how it pans out for the mobility sector.

Author

Vikram Kothamachu : Senior Technical Lead- VED Practice Group, KPIT

Abhilash SK : Sr Technical Lead (Deep learning , Computer vision , Machine learning ,Image processing), PathPartner Technology