Sim-to-real and the Isaac stack

10 atomic recall cards. Export to Anki and let spaced repetition do its slow work.

In Anki: File → Import, choose this TSV, set field separator to Tab, deck = Tokens to Embodied Minds · Ch 29, note type = Basic.

Front	Back
What is domain randomization?	Training a policy across a distribution of simulator parameters (friction, mass, lighting, etc.) so that the real world is one sample from that distribution and the policy generalizes to it without ever seeing it.
What is system identification in the context of sim-to-real?	Measuring real robot dynamics (friction, damping, motor curves, sensor noise) and fitting simulator parameters to match, centering the domain randomization distribution on real-world values.
What is the Newton physics engine and when was it announced?	A new GPU-accelerated physics engine built in NVIDIA Warp, co-developed with DeepMind and Disney Research. Announced at NVIDIA GTC, March 18, 2025. Delivers ~70x simulation throughput vs previous stacks.
What is Isaac Lab?	The robot learning framework built on Isaac Sim (NVIDIA Omniverse): task definitions, reward functions, domain randomization APIs, and PPO/SAC training loops. The open stack for sim-to-real robot ML.
How many hours of synthetic data did GR00T-Dreams generate in 11 hours?	6,500 hours of synthetic humanoid training data, from a real-seed demonstration dataset, using Isaac Lab and Newton.
What improvement did GR00T N1.5 show on RoboCasa 30-demo vs N1?	From 17.4% to 47.5% task success — driven largely by DreamGen synthetic data generated via GR00T-Dreams in Isaac Lab.
What is the difference between Isaac Sim and Isaac Lab?	Isaac Sim: photorealistic simulation environment (NVIDIA Omniverse). Isaac Lab: robot learning framework built on Isaac Sim — task APIs, reward functions, domain randomization, training loops.
Name three physics parameters commonly randomized in Isaac Lab for a manipulation task.	Friction coefficient of the table surface, mass of the object being grasped, joint damping of the robot arm.
What is the primary reference for domain randomization?	Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World, Tobin et al., arXiv:1703.06907, March 20, 2017.
How does GR00T-Dreams generate synthetic data from real demonstrations?	It retargets real demonstration videos to the sim environment using video-to-motion retargeting and scene reconstruction, then augments with Isaac Lab rollouts using domain randomization — multiplying the real seed data into thousands of synthetic hours.