LingBot-Video, developed by @robbyant_brain, has introduced a groundbreaking open-source video generation model that leverages a Mixture-of-Experts (MoE) design to enhance performance in embodied AI scenarios. This model, which activates only 3 billion parameters during generation out of its flagship 30 billion, achieves nearly three times the inference efficiency compared to traditional dense architectures. It stands out as it is trained not just on standard web videos but incorporates over 70,000 hours of embodied footage, which includes examples of robot manipulation and first-person actions, ensuring it learns real-world dynamics effectively. Such capabilities align with the broader trend in robotics where video models are used for world simulation, policy evaluation, and action planning.

RBench: RBench is a benchmark designed to evaluate video generation models on metrics including spatial understanding, motion coherence, and embodiment-related tasks. LingBot-Video achieved a leading average score on RBench compared to other open and closed models listed in the evaluation. The benchmark highlights performance in areas such as physical interactions and object permanence in generated videos.
LingBot-Video: LingBot-Video is an open-source Diffusion Transformer video foundation model designed for embodied AI scenarios and built with a Mixture-of-Experts architecture. Developed by Robbyant Brain, it integrates specialized pretraining on embodied footage with web videos to emphasize physical plausibility, causality, and action consequences in generated outputs. The model supports robotics use cases by functioning as a world simulator for predicting future frames from initial states and action sequences.
Robbyant Brain: Robbyant Brain is the organization responsible for creating and releasing LingBot-Video, including its open-source code, model weights, and accompanying research paper. It focuses on advancing video generation systems that incorporate real-world physical interactions rather than relying solely on internet video data. The entity has made the model available through Hugging Face and GitHub to support further research in embodied AI.

Robotics Applications: Video models trained with action-to-video capabilities are increasingly used as world simulators for data generation, policy evaluation, and action planning in robotics.
Embodied Training Data: Video foundation models benefit from mixing embodied footage such as robot manipulation and egocentric action clips with standard web videos to better learn real-world dynamics and consequences.
Mixture-of-Experts Efficiency: Sparse Mixture-of-Experts architectures in video models enable higher total parameter capacity while activating only a fraction of parameters per generation, supporting longer sequence handling.