AI development of deep learning models is an iterative process that requires a data centric approach. Synthetic data generation is a low cost method to accelerate the iterative process of model development. Developers can target corner/test cases, and conditions that are otherwise impossible to collect in the real world due to cost or safety with synthetic data.
We announced Omniverse Replicator which provides a platform for synthetic data generation. We released the Isaac SIM Replicator Composer application powered by Omniverse this month, in source for data developers to use for synthetic data generation.
Omniverse Replicator for Isaac Sim
This application was developed as the digital cockpit for our data developers. We use Replicator Composer to create datasets for training deep learning perception functions for robotics applications. The application takes a production flow data centric approach with control of parameters in YAML, and tracing of tools and parameters used in each dataset.
The Isaac Sim release includes more than 100 3D models, textures and materials to start development of datasets with domain randomization. Replicator Composer includes example YAML inputs and asset lists to get started
Trained models from the synthetically generated data can be deployed with the Isaac ROS DNN Inference package.
Isaac Sim is available for download to get started generating your own synthetic datasets for machine learning.