Shields to Guarantee Probabilistic Safety in MDPs for Autonomous Agents

arXiv·medium signal

Extends classical shielding from absolute safety (nothing bad ever happens) to probabilistic safety (bad events below acceptable probability threshold) for autonomous agents. The formal framework conservatively extends classical shields while maintaining maximal permissiveness — agents can take useful actions that deterministic shields would block, as long as cumulative risk stays below the budget. Relevant to deploying agents in stochastic environments.