Denso and Carnegie Mellon University (CMU) will present new research focused on 4D driving scene generation for autonomous vehicle simulation at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026.
The research was developed through a collaboration between Denso’s Pittsburgh Innovation Lab and the Robotics Institute at Carnegie Mellon University, with the goal of improving the generation of synthetic training data for autonomous driving systems.
New Approach to Autonomous Vehicle Simulation
The research introduces a grounded latent representation framework that models vehicles, pedestrians, and environmental elements as separate, editable components within a driving scene.
By representing each scene as an entity-centric collection of latent variables, the system enables precise control over individual objects, their behavior, and movement patterns. According to the researchers, this approach allows the creation of dynamic driving scenarios while maintaining stable backgrounds and realistic object trajectories.
The technology is intended to improve simulation environments used for training and validating autonomous vehicle systems, particularly in rare or hazardous edge cases that are difficult to capture through real-world data collection.
Focus on Scalable Synthetic Data Generation
As autonomous driving developers increasingly rely on simulation to accelerate development, the ability to generate high-quality synthetic data has become a critical challenge.
The Denso-CMU method aims to support more scalable simulation workflows by enabling the generation of diverse driving scenarios while preserving physical realism and environmental consistency.
The researchers believe the approach can help improve virtual testing, scenario generation, and AI model training for future autonomous vehicle systems.
Collaboration Between Industry and Academia
The project highlights ongoing collaboration between industry and academic institutions in advancing artificial intelligence technologies for transportation applications.
“Our collaboration with DENSO demonstrates how academic research and industry expertise can come together to solve complex challenges in AI and perception,” said Kris Kitani, Associate Research Professor at the Robotics Institute within Carnegie Mellon University’s School of Computer Science.
“By advancing capabilities such as simulation, model efficiency and sensing, this work helps accelerate the path from research to real-world autonomous driving applications.”
Supporting the Full Autonomous Driving Development Stack
Denso says the research aligns with its broader strategy of advancing technologies across the entire autonomous driving development process, including model training, simulation, validation, and deployment.
“This research with Carnegie Mellon University is a vital step forward in scaling AI world models for autonomous driving,” said Shawn Hunt, Software Engineer at Denso’s Pittsburgh Innovation Lab.
“It opens the door to more reliable virtual testing, scenario generation and training environments for future autonomous vehicle systems – all of which translates to safer and more effective vehicle autonomy.”
Toru Hirano, Vice President of North America Research and Development at Denso, said the project reflects the company’s continued focus on strengthening the AI development stack required for future automated driving technologies.
Additional Research To Be Presented
Alongside the Pittsburgh Innovation Lab project, Denso will also present three additional academic papers developed by its Denso IT Laboratory in Japan at CVPR 2026.
The company noted that these papers are separate research contributions and are not part of the Carnegie Mellon University collaboration.
CVPR is one of the world’s leading conferences in computer vision and artificial intelligence, bringing together researchers and technology companies to showcase advancements in machine learning, perception systems, robotics, and autonomous mobility.