世界模型ICLR 20262026
World Model Agents: Scaling Embodied Intelligence
Google DeepMindGoogle DeepMind
摘要
We present World Model Agents, a framework for scaling embodied intelligence through world models. Our agents learn rich internal world representations from diverse experience, enabling zero-shot transfer to novel tasks and environments with human-level performance across hundreds of benchmarks.
embodied intelligenceworld model agentzero-shot transferscalable learninghuman-level performance
技术细节
数据集
测试机器人
模型骨架
VLA (Vision-Language-Action) Transformer
编码器
ViT-L/14 (视觉), LLaMA-2 (语言)
解码器
Diffusion Policy Action Head
论文深度解读
阅读完整解读本文包含详细的论文深度解读，涵盖模型架构、核心公式、数据策略、实验解剖等内容。