SANTA CLARA, Calif. — NVIDIA is introducing a data generation engine to produce physically simulated synthetic data for training artificial learning (AI) deep neural networks.
NVIDIA Omniverse Replicator augments costly and laborious human-labeled real-world data, which can be error prone and incomplete, with the ability to create large and diverse physically accurate data, according to the company last month.
It also enables the generation of ground-truth data that is difficult or impossible for humans to label, such as velocity, depth, occluded objects, adverse weather conditions or tracking the movement of objects across sensors.
NVIDIA introduced two applications for generating synthetic data: one for NVIDIA DRIVE Sim, a virtual world for hosting the digital twin of autonomous vehicles; and one for NVIDIA Isaac Sim, a virtual world for the digital twin of manipulation robots.
Omniverse Replicator will be made available next year to developers to build domain-specific data generation engines.
The replicator allows developers to bootstrap AI models, fill real-world data gaps, and label the ground truth. Data generated in the virtual worlds can cover a range of scenarios, including rare or dangerous conditions that can’t regularly or safely be experienced in the real world.
Autonomous vehicles and robots built using the data can “master skills” across virtual environments “before applying them in the physical world.”
Omniverse Replicator allows users to “create diverse, massive, accurate data sets to build high-quality, high-performing and safe data sets, which is essential for AI,” said Rev Lebaredian, VP of simulation technology and Omniverse engineering, NVIDIA.
“While we have built two domain-specific data generation engines ourselves, we can imagine many companies building their own with Omniverse Replicator,” Lebaredian said.
Omniverse Replicator is part of NVIDIA Omniverse, a virtual world simulation and collaboration platform for 3D workflows.