arXiv:2609.35874 Applies CVaR to Per-Step Belief Costs in POMDPs

The formulation computes CVaR_alpha of the state-dependent cost under the current belief instead of its expectation. This isolates risk inside the belief without altering the value function. Existing planners require only a changed cost oracle. Finite-time bounds on policy evaluation and sparse sampling hold with error independent of alpha. A new particle-belief MDP approximation gap bound closes the chain from true POMDP value to algorithmic output.

Standard POMDP solvers such as SARSOP and DESPOT optimize expectation and therefore ignore mass on high-cost states when belief variance is large. Earlier CVaR work on trajectory returns required bespoke algorithms. The new per-step CVaR restores MDP structure, recovering risk-neutral planning exactly at alpha = 1. End-to-end guarantees therefore transfer without modification.

Deployment of learned perception pipelines in robotics and autonomous systems produces beliefs whose tails correspond to real collision or failure modes. The method supplies a tunable knob that contracts those tails while preserving planner codebases already in production. No new solver is required.

The paper leaves open scaling to continuous observation spaces and online belief compression under CVaR. Follow-on work will likely test the bound tightness on standard benchmarks such as RockSample and Tag.