This six-page perspective argues that high likelihood and visual fidelity do not ensure a world model keeps the evidence needed for safe decisions. It names three mismatches (likelihood versus risk, prediction versus intervention, finite-horizon prediction versus accumulated consequences) and proposes the Risk-Informed World Model, organised around consequences, intervention, epistemic uncertainty and recoverability. Its four parts are decision-relevant representation, counterfactual reasoning, safety-critical episodic memory and runtime safety assurance.
Relevance
Recoverability and revisable safety memory are framed here as world-model properties, which ties long-term autonomy to robust robotics. A position paper with no experiments, so read it for the questions; see safe control.
Links
- arXiv
- Subfield: Safety and verification, Semantic mapping, Perception