X-Humanoid: Robotize Human Videos to Generate Humanoid Videos at Scale

#XHumanoid #Robotics #EmbodiedAI #DiffusionTransformer #DataGeneration #TeslaOptimus #humanoid #tending #technology X-Humanoid is a groundbreaking generative video editing framework developed to solve the critical problem of data scarcity that hinders the progress of embodied AI, including Vision-Language-Action (VLA) models and world models for humanoid robots. By transforming vast existing human activity videos into robot footage, X-Humanoid aims to bridge the **visual embodiment gap** that previously prevented the direct use of human data for training robot policies. **The Generative Video Editing Approach:** At its core, X-Humanoid introduces a generative video editing approach that adapts the powerful **Wan 2.2 Diffusion Transformer (DiT)** video generation model into a novel **video-in video-out structure**. The finetuned model accepts a third-person human video and outputs a corresponding video where the person is seamlessly replaced by a specific humanoid embodiment, such as the **Tesla Optimus humanoid**. A critical constraint enforced by the model is that the humanoid's motion must be **strictly synchronised** with the original human's action, while preserving all other video content, such as the background scene. **Technical Architecture Details:** To achieve this video translation, the input human video is encoded into non-denoising **condition tokens**. These tokens are concatenated with the denoising **generation tokens**. During the self-attention process within the DiT blocks, a mask is applied to ensure the condition tokens do not attend to the generation tokens, preventing corruption of the input video data. The model is finetuned using a flow-matching objective, learning to predict the velocity of a path between noise and the clean latent tokens, ultimately decoding only the generation tokens to produce the edited video. **Scalable Data Synthesis:** Training this video-to-video translation model required paired human-humanoid videos, which w