Dataset · nvidia
NVIDIA synthetic robot scene videos
nvidia/PhysicalAI-WorldModel-Synthetic-Embodied-Robot-Scenes
Large corpus of synthetic videos with robots from NVIDIA simulators for training physical world models.
What's inside
The dataset from NVIDIA includes between 100,000 and 1 million synthetic video recordings generated in simulation and rendering pipelines based on USD (NVIDIA Isaac Sim, Omniverse, Isaac Lab). The data is created to improve physical plausibility, maintain embodiment persistence, and understand robot behavior during task execution.
What you can do with it
- Training world models to predict physics and robot behavior
- Developing and fine-tuning VLA (Vision-Language-Action) models and robot control systems
- Researching the transfer of robot skills from simulation to the real world (Sim2Real)
- Size
- 100K–1M rows
- Downloads this month
- 3.9K
- Likes
- 41
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/PhysicalAI-WorldModel-Synthetic-Embodied-Robot-Scenes")