GTC Silicon Valley-2019 ID:S9596:Simulation to Reality Transfer in Robotic Learning
The growing prevalence of synthetic data for training deep learning networks will revolutionize robotic systems. Synthetic data holds the promise of nearly unlimited pre-labeled training data that is generated safely out of harm's way. One of the key challenges involved in using synthetic data is bridging the reality gap to ensure that networks trained on this data operate correctly when exposed to data from the real world. We'll discuss our recent progress in exploring sim-to-real transfer in robotics and explore the reality gap for real-time detection and pose estimation of known household objects from a single RGB image. We'll show that for this problem the reality gap can be spanned by a simple combination of domain randomized and photorealistic data. We will also discuss the use of reinforcement learning for sim-to-real robotic problems such as grasping, peg-in-hole insertion, and articulated interaction.