World ModelsarXiv preprint2025
Learning Massively Multitask World Models for Continuous Control
Abhinav Shaw, Natcha Simsiri, Iman Deznaby, Madalina Fiterau, Tauhidur Rahaman et al.Tsinghua University, BAAI
摘要
We present a framework for learning massively multitask world models that generalize across diverse continuous control tasks. Our approach trains a single world model on hundreds of different environments simultaneously, enabling strong zero-shot and few-shot transfer to novel tasks. The model learns shared dynamics representations while maintaining task-specific adaptations, achieving state-of-the-art results on multiple continuous control benchmarks.
multitaskworld modelcontinuous controlmassive learninggeneralization
技术细节
数据集
DMC SuiteMeta-WorldProcgen
仿真平台
MuJoCoDM Control
模型骨架
Multitask Transformer World Model
编码器
State Encoder + Task Embedding
解码器
Dynamics Decoder + Action Head
