World Model for Robot Learning: A Comprehensive Survey

World models—predictive representations of how environments evolve under actions—are becoming central to robot learning. Now, a landmark survey by 18 top researchers including Philip Torr, Pieter Abbeel, Jitendra Malik, and others provides a comprehensive review of this rapidly growing field. The 43-page paper systematically examines how world models are coupled with robot policies, serve as learned simulators for reinforcement learning, and enable evaluation and planning.

The survey covers the progression from imagination-based generation to controllable, structured, and foundation-scale formulations, especially in robotic video generation. It connects these ideas to practical domains like navigation and autonomous driving, summarizing key datasets, benchmarks, and evaluation protocols. The authors aim to unify fragmented literature across architectures, functional roles, and embodied application domains.

This work highlights major challenges and future directions for predictive modeling in embodied agents. The accompanying GitHub repository will be updated regularly to track new works and resources. For AI professionals, this survey offers a definitive roadmap for building robots that can anticipate and adapt to their environment—a critical step toward general-purpose robotics.