Sber has introduced the Kandinsky WM 1.0 family of generative models, designed to create video data with precise adherence to the laws of physics. The development is aimed at training Physical AI systems – artificial intelligence that controls robots, autonomous vehicles, and industrial equipment.
Neural networks solve the problem of training data scarcity: modern models require millions of hours of video recordings for training, and collecting such materials is extremely difficult and expensive. Kandinsky WM allows generating rare and dangerous scenarios – accidents, extreme weather conditions, equipment failures – which are impossible to film in real conditions.
The models are based on Kandinsky 5.0 Video Lite, fine-tuned on several million video recordings from robot cameras, autonomous vehicles, and industrial facilities. Special attention was paid to physical plausibility: ensuring that objects did not deform, disappear, or move unnaturally. For automotive scenarios, reinforcement learning was applied: special models evaluated the quality of the videos and provided feedback.
The models generate videos lasting about five seconds, reflecting individual actions: manipulating objects, road maneuvers, stages of production processes. Kandinsky WM 1.0 is already being used by Sber's Robotics Center to train robots and by the AIRI institute for the development of a road simulator.
The code and model weights are published under the open MIT license – developers worldwide can use them for free. As the developers note, Kandinsky WM is the first step towards creating full-fledged world models capable of simulating scene evolution over time and serving as simulators for robot training.