Introducing LingBot-Video: A Large-Scale Mixture-of-Experts Model for Embodied Intelligence Pretraining

Researchers have developed LingBot-Video, a groundbreaking video pretraining paradigm designed for embodied intelligence, addressing the limitations of existing video generative models. Unlike traditional models that prioritize visual fidelity, LingBot-Video employs a Mixture-of-Experts architecture to enhance modeling capacity while improving inference efficiency. This model is trained using a unique data profiling engine that integrates standard internet videos with robot-specific footage, enabling a deeper understanding of manipulation, navigation, and action dynamics. Additionally, a multi-dimensional reward system has been implemented to ensure alignment with physical rationality and task completion. Comprehensive evaluations demonstrate the model's performance and efficiency, marking it as the first large-scale, open-source MoE video foundation model aimed at connecting digital creativity with physical actions.