# MBZUAI’s KineMask tests video models’ response to object collisions > The method gives a model an object’s initial velocity and asks it to infer what motion follows. - Canonical URL: https://newsverum.com/en/news/mbzuai-kinemask-video-physical-interactions-2026 - Publisher: Newsverum - Section: Technology - Country: United Arab Emirates - Published: 2026-10-20T14:20:10+00:00 - Updated: 2026-10-20T14:20:10+00:00 - Event / data period: 2026-10-12 - Other editions: [Türkçe](https://newsverum.com/tr/news/mbzuai-kinemask-video-physical-interactions-2026.md), [العربية](https://newsverum.com/ar/news/mbzuai-kinemask-video-physical-interactions-2026.md), [Français](https://newsverum.com/fr/news/mbzuai-kinemask-video-physical-interactions-2026.md), [Español](https://newsverum.com/es/news/mbzuai-kinemask-video-physical-interactions-2026.md), [Deutsch](https://newsverum.com/de/news/mbzuai-kinemask-video-physical-interactions-2026.md), [Português](https://newsverum.com/pt/news/mbzuai-kinemask-video-physical-interactions-2026.md) MBZUAI researchers developed KineMask, a method that gives a video-generation model an object’s starting direction and speed, then asks it to predict the motion that follows. A user selects an object in an image and specifies an initial motion; the model generates a scene that may include collisions and their consequences. The approach provides an initial condition rather than a full frame-by-frame trajectory. In a synthetic Interactions test described in MBZUAI’s 12 October technical blog, CogVideoX recorded an FVD score of 807.3, while KineMask trained on interaction data scored 250.7. Their Video Motion Distance scores were 3,514.9 and 143.8, respectively. The results are limited to the synthetic benchmark and the researchers’ comparison; the study does not establish that the system has learned physical laws internally. ## Data ### KineMask synthetic-interaction benchmark Comparison on a synthetic-interaction dataset Methodology: The unit registry has no matching units for FVD or Video Motion Distance scores; figures remain in the article text with model and synthetic-dataset context. FVD and Video Motion Distance results are limited to the named models and synthetic-dataset conditions; no cross-model or cross-country comparison or aggregation is made. | date | label | description | status | |---|---|---|---| | 2026-10-12 | Research findings published | The study reports model results under the synthetic or controlled benchmark conditions stated in the source. | occurred | Sources: · Download data: [JSON](https://newsverum.com/api/v1/articles/mbzuai-kinemask-video-physical-interactions-2026/data/synthetic-interactions-benchmark?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/mbzuai-kinemask-video-physical-interactions-2026/data/synthetic-interactions-benchmark?lang=en&format=csv) ## Sources - [Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)]() (Source checked: 2026-10-20; Published: 2026-10-12) --- Cite the canonical URL above. [Editorial standards and corrections](https://newsverum.com/en/editorial)