Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence

2026-07-08

AI Summary

Existing video generation models predominantly emphasize content creation while neglecting computational efficiency and physical plausibility, rendering them ill-suited for embodied intelligence applications. This work proposes LingBot-Video, a DiT-based video pretraining paradigm tailored for embodied intelligence, which introduces the first large-scale open-source Mixture-of-Experts (MoE) video foundation model. The model integrates diverse data modalities—including robotic manipulation, navigation, and first-person perspective videos—and incorporates a multidimensional reward mechanism aligned with physical plausibility and task completion. Experimental results demonstrate that LingBot-Video significantly enhances both the physical realism and task-oriented fidelity of generated videos while maintaining efficient inference, thereby effectively bridging the gap between digital generation and real-world physical execution.

Abstract

Despite the recent promise in robot control, video generative models suffer from a domain mismatch due to their primary focus on content creation. For example, their design inherently prioritizes visual fidelity and creativity over computational efficiency and physical realism. In this work, we present LingBot-Video, a DiT-based video pretraining paradigm specifically tailored for embodied intelligence. From the architecture perspective, we adopt the Mixture-of-Experts (MoE), instead of dense, framework to achieve a better trade-off between modeling capacity and inference efficiency, and manage to scale it up from scratch. From the data perspective, we construct a data profiling engine that augments standard internet videos with extensive robot-oriented footage, encompassing manipulation, navigation, and egocentric perspectives, to equip the base model with an intrinsic understanding of actions and world dynamics. From the training perspective, we develop a multi-dimensional reward system to enforce the alignment regarding physical rationality and task completion, going beyond standard criteria such as aesthetics, prompt-following, and motion consistency. Comprehensive evaluations validate its performance and efficiency as a video foundation model. We contribute LingBot-Video as the inaugural large-scale, open-source MoE video foundation model to the community, in a pioneering effort to bridge digital creativity and physical actuation.

Problem

• embodied intelligence
• video generative models
• domain mismatch
• physical realism
• computational efficiency

Innovation

• Mixture-of-Experts
• video pretraining
• embodied intelligence
• DiT architecture
• physical rationality