ShengShu Technology launches Vidar multi-view physical AI training model - The Robot Report

ShengShu Technology, a Beijing-based company founded in March 2023 specializing in multimodal large language models, has launched Vidar, a multi-view physical AI training model designed to accelerate robot development. Vidar, which stands for “video diffusion for action reasoning,” leverages a combination of limited physical training data and generative video simulations to train embodied AI models. Unlike traditional methods that rely heavily on costly, hardware-dependent physical data collection or purely simulated environments lacking real-world variability, Vidar creates lifelike multi-view virtual training environments. This approach allows for scalable, robust training of AI agents capable of real-world tasks, reducing the need for extensive physical data by up to 1/80 to 1/1,200 compared to industry-leading models. Built on ShengShu’s flagship video-generation platform Vidu, Vidar employs a modular two-stage learning architecture that separates perceptual understanding from motor control. In the first stage, large-scale general and embodied video data train the perceptual