Reliability-aware dispatch for autonomous fleets: A reproducible risk-to-policy pipeline for reducing failures and downtime

Autonomous mobility fleets operate under tight service-level constraints while facing mission-dependent failure risk and nontrivial downtime costs. Traditional dispatch optimizes demand coverage (or revenue) and treats reliability as an exogenous maintenance process; this can systematically allocate high-stress missions to already risky vehicles, increasing roadside failures and service disruption. We present a risk-to-policy pipeline that connects a mission-level failure risk model to actionable fleet dispatch decisions. The pipeline produces per-mission predicted failure probability, calibrates it to observed outcomes, and exposes an operating point (risk threshold / ranking rule) that can be tuned to trade off fleet throughput against failure and downtime. We evaluate two policies on multi-day traces: a demand-driven BASELINE and a REL-AWARE policy that ranks candidate vehicles by predicted risk, routes high-risk vehicles to preventive maintenance, and reserves low-risk vehicles for longer or higher-value missions. Across days, REL-AWARE improves failure capture (lift) and yields measurable reductions in mission failures with comparable mission service, while providing an interpretable control knob for operators. The full workflow is reproducible and designed to plug into existing dispatch stacks.