Generate Synthetic Data of Dynamic Environments With a Custom Isaac Sim-Based Framework
, Ph.D. Student, Max Planck Institute for Intelligent Systems
Robots need to be simulated before they can be deployed in the real world. Unfortunately, this is usually done in static scenarios even though the real world is dynamic. High-quality simulations of dynamic environments will help us build robots that can be safely deployed in the real world. These simulations need to be photorealistic and physically accurate, with sensor noise, and be flexible enough to adapt to different sensors and setups. However, no system currently works like that out of the box. We'll show how this can be done using NVIDIA Isaac Sim, publicly available datasets, Omniverse Connectors, and our framework to generate data with a simplified learning curve. The five key ingredients include (1) environments, (2) dynamic humans, (3) simulation setup and placement of the assets, (4) simplified robot control, and (5) data saving and post-processing.
It's not required, but familiarity with Isaac Sim and Python will help you better understand this talk.