arXiv:2512.04040v1 Announce Type: new 
Abstract: A truly interactive world model requires three key ingredients: real-time long-horizon streaming, consistent spatial memory, and precise user control. However, most existing approaches address only one of these aspects in isolation, as achieving all three simultaneously is highly challenging-for example, long-term memory mechanisms often degrade real-time performance. In this work, we present RELIC, a unified framework that tackles these three challenges altogether. Given a single image and a text description, RELIC enables memory-aware, long-duration exploration of arbitrary scenes in real time. Built upon recent autoregressive video-diffusion distillation techniques, our model represents long-horizon memory using highly compressed historical latent tokens encoded with both relative actions and absolute camera poses within the KV cache. This compact, camera-aware memory structure supports implicit 3D-consistent content retrieval and enforces long-term coherence with minimal computational overhead. In parallel, we fine-tune a bidirectional teacher video model to generate sequences beyond its original 5-second training horizon, and transform it into a causal student generator using a new memory-efficient self-forcing paradigm that enables full-context distillation over long-duration teacher as well as long student self-rollouts. Implemented as a 14B-parameter model and trained on a curated Unreal Engine-rendered dataset, RELIC achieves real-time generation at 16 FPS while demonstrating more accurate action following, more stable long-horizon streaming, and more robust spatial-memory retrieval compared with prior work. These capabilities establish RELIC as a strong foundation for the next generation of interactive world modeling.

تم تقديم RELIC كنموذج مبتكر لعالم تفاعلي يدمج البث المباشر على المدى الطويل، والذاكرة المكانية المتسقة، والتحكم الدقيق من قبل المستخدم، مما يعالج التحديات التي تواجه النماذج الحالية لتحقيق هذه الوظائف في وقت واحد.

RELIC se ha presentado como un modelo de mundo interactivo innovador que integra la transmisión en tiempo real a largo plazo, una memoria espacial coherente y un control preciso del usuario, abordando los desafíos que enfrentan los modelos existentes para lograr estas funcionalidades simultáneamente.

RELIC a été présenté comme un modèle de monde interactif innovant qui intègre le streaming en temps réel à long terme, une mémoire spatiale cohérente et un contrôle utilisateur précis, répondant aux défis rencontrés par les modèles existants pour atteindre ces fonctionnalités simultanément.

RELIC has been introduced as an innovative interactive world model that integrates real-time long-horizon streaming, consistent spatial memory, and precise user control, addressing the challenges faced by existing models in achieving these functionalities simultaneously.

RELIC: Interactive Video World Model with Long-Horizon Memory

arXiv:2512.04537v1 Announce Type: new 
Abstract: The advancement of embodied AI has unlocked significant potential for intelligent humanoid robots. However, progress in both Vision-Language-Action (VLA) models and world models is severely hampered by the scarcity of large-scale, diverse training data. A promising solution is to "robotize" web-scale human videos, which has been proven effective for policy training. However, these solutions mainly "overlay" robot arms to egocentric videos, which cannot handle complex full-body motions and scene occlusions in third-person videos, making them unsuitable for robotizing humans. To bridge this gap, we introduce X-Humanoid, a generative video editing approach that adapts the powerful Wan 2.2 model into a video-to-video structure and finetunes it for the human-to-humanoid translation task. This finetuning requires paired human-humanoid videos, so we designed a scalable data creation pipeline, turning community assets into 17+ hours of paired synthetic videos using Unreal Engine. We then apply our trained model to 60 hours of the Ego-Exo4D videos, generating and releasing a new large-scale dataset of over 3.6 million "robotized" humanoid video frames. Quantitative analysis and user studies confirm our method's superiority over existing baselines: 69% of users rated it best for motion consistency, and 62.1% for embodiment correctness.

يمثل تقديم X-Humanoid تقدمًا كبيرًا في مجال الذكاء الاصطناعي المجسد، مما يتيح تحويل مقاطع الفيديو البشرية إلى تمثيلات إنسانية على نطاق واسع. تستخدم هذه الطريقة في تحرير الفيديو التوليدي نموذج Wan 2.2، حيث يتم تكييفه لمهمة ترجمة الإنسان إلى إنسان آلي، وهو أمر حاسم لتدريب الروبوتات الذكية. تم إنشاء خط أنابيب لإنشاء البيانات القابلة للتوسع لتوليد أكثر من 17 ساعة من مقاطع الفيديو المتزاوجة بين الإنسان والإنسان الآلي، مما يلبي الحاجة إلى بيانات تدريب متنوعة.

La introducción de X-Humanoid marca un avance significativo en el campo de la IA encarnada, permitiendo la transformación de videos humanos en representaciones humanoides a gran escala. Este enfoque de edición de video generativo utiliza el modelo Wan 2.2, adaptándolo para la tarea de traducción de humano a humanoide, lo cual es crucial para el entrenamiento de robots inteligentes. Se ha establecido un pipeline de creación de datos escalable para generar más de 17 horas de videos emparejados de humano-humanoide, abordando la necesidad de datos de entrenamiento diversos.

L'introduction de X-Humanoid représente une avancée significative dans le domaine de l'IA incarnée, permettant la transformation de vidéos humaines en représentations humanoïdes à grande échelle. Cette approche de montage vidéo génératif utilise le modèle Wan 2.2, l'adaptant pour la tâche de traduction humain-humanoïde, ce qui est crucial pour former des robots intelligents. Un pipeline de création de données évolutif a été établi pour générer plus de 17 heures de vidéos humaines-humanoïdes appariées, répondant ainsi au besoin de données d'entraînement diversifiées.

The introduction of X-Humanoid marks a significant advancement in the field of embodied AI, enabling the transformation of human videos into humanoid representations at scale. This generative video editing approach utilizes the Wan 2.2 model, adapting it for the human-to-humanoid translation task, which is crucial for training intelligent robots. A scalable data creation pipeline has been established to generate over 17 hours of paired human-humanoid videos, addressing the need for diverse training data.

RELIC: Interactive Video World Model with Long-Horizon Memory

Was this article worth reading? Share it

LucidQuery AI

AIvilization

Mixhubb