World Model for Robot Learning: A Comprehensive Survey

Authors: Bohan Hou, Gen Li, Jindou Jia, Tuo An, et al. (including Pieter Abbeel, Jitendra Malik, Jianfei Yang)

arXiv: 2605.00080 (2026)

Abstract

Reviews world models — predictive representations of how environments evolve under actions — as central components in robot learning. Covers integration with robot policies, use as learned simulators for RL, and progression from imagination-based generation to foundation-scale formulations.

Scope

• Policy learning with world models (model-based RL, imagination rollouts)
• World models as simulators: offline policy evaluation, OOD testing, safety probing
• Robotic video generation and data synthesis
• Navigation, manipulation, autonomous driving applications
• Foundation-scale world models (Genie, DIAMOND, IRASim)

Safety-Relevant Contributions

• Offline policy evaluation: use world model to evaluate a policy before real-world deployment
• OOD testing: probe policy behavior under edge cases generated by the world model
• Safety probing: generate adversarial scenarios to find failure modes before deployment

Notes

One of the most comprehensive recent surveys (2026). Safety is treated as an application of simulation capability rather than a first-class concern — the framing is “world models enable safer development” rather than “world models have safety properties.”