ubr-physical-ai2026cc-by-4.0

Cosmos3 image-to-video: subject survival on synthetic rescue scenes

26 generated video clips from NVIDIA Cosmos3 models conditioned on synthetic Isaac Sim frames, with per-frame detection scores measuring whether image-to-video generation maintains subject presence across frames. Published as a negative result demonstrating that image-conditioned generation fails to hold subjects in place without per-frame control signals.

Downloads0
Episodes26

Why This Matters for Physical AI

Demonstrates fundamental limitations of image-conditioned video generation for maintaining object persistence in synthetic scenes, informing design choices for perception models in search-and-rescue robotics applications.

Technical Profile

Modalities
rgbvideo
Environment
simulation
Task Types
search-and-rescue
Episodes
26
Data Format
mp4
Annotation Types
bounding_boxesdetection_labels
License
cc-by-4.0
Part of the NVIDIA Cosmos3 family

Access

Need custom rgb data?

Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets