PhysicalAI-WorldModel-Synthetic-Physical-Interaction-Scenes

PhysicalAI-WorldModel-Synthetic-Physical-Interaction-Scenes Dataset Card Dataset Description PhysicalAI-WorldModel-Synthetic-Physical-Interaction-Scenes is a large-scale synthetic dataset of physically-simulated multi-object interaction scenes, generated using NVIDIA Isaac Sim and the PhysX physics engine. It is designed to train and evaluate AI models on physical reasoning, rigid body dynamics, optical flow, depth estimation, and scene understanding. Each clip is a 5–8 second, 30 FPS, 1920×1080 simulation rendered from one or more fixed camera viewpoints, paired with ground-truth physics annotations including per-object velocity, angular velocity, center-of-mass displacement, cumulative rotation, per-frame semantic segmentation, lossless depth video, and VLM-generated natural language captions grounded in the underlying physics data. Most scenes produce 5-second clips (150 frames); the `ball_mixer` scene produces 8-second clips (240 frames). The dataset contains ten procedurally-parameterized scene types (`dominoes`, `ball_mixer`, `bowling`, `billiards`, `towers`, `wrecking_ball`, `objects_falling`, `rolling_ramp_objects`, `rolling_ramp_obstruct`, `obstruction`), each exercising different physical phenomena such as cascading impact chains, multi-body collisions, rotational energy transfer, and ballistic trajectories. This dataset is ready for commercial/non-commercial use. This dataset is fully synthetic. It contains no real-world imagery, no personally identifiable information, and no individuals. All content was generated procedurally using NVIDIA Isaac Sim with open or NVIDIA-licensed 3D assets. This dataset is ready for commercial or non-commercial uses. Dataset Owner(s) NVIDIA Corporation Dataset Creation Date May 2026 Version.** [1.0] License/Terms of Use This dataset is released under the OpenMDW1.1 Intended Usage This dataset is intended for researchers and AI practitioners working on: **Physical reasoning** from video — predicting object trajectories, coll