Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence

This paper introduces LingBot-Video, a DiT-based video pretraining paradigm designed for embodied intelligence rather than primarily for visual content creation. It uses a Mixture-of-Experts architecture to increase modeling capacity while improving inference efficiency, and scales the model from scratch. A data profiling engine augments internet video with robot-oriented footage covering manipulation, navigation, and egocentric perspectives. Training adds a multidimensional reward system targeting physical rationality and task completion, beyond aesthetics, prompt following, and motion consistency. The authors describe it as the first large-scale open-source MoE video foundation model aimed at connecting visual generation with physical actuation.