Skip to main content

Fleet Intelligence

The fleet-intelligence cognition layer (T5 — Operate): fleet-wide telemetry collection, operational dashboards, drift detection, and retraining triggers across a fleet of deployed robots. "Fleet" means a fleet of robots, not Kubernetes clusters. This is the cognition layer that sits above the implemented fleet-delivery control plane (T4).

[!WARNING] Roadmap / placeholder, not shipped. The fleet-intelligence domain ships 0 Python files and design specifications only (4 placeholder specs). Everything below describes intended capability, not working code. Treat this domain as a roadmap direction, not an available feature. For the canonical tier model, the autonomy ladder, and fleet vocabulary, see tier-model.md.

Human-in-the-loop is the default; closed-loop retraining is a foot-gun. Fully autonomous retraining on production data can bake current degraded behavior into the next dataset, and drift detection needs statistical power that only exists at fleet scale. Fleet intelligence should default to human-supervised stages (T5.0–T5.1), not the autonomous closed loop (T5.3). See the autonomy ladder.

📋 Prerequisites

RequirementPurpose
Azure IoT OperationsEdge telemetry collection and MQTT brokering
Azure Event HubsCloud telemetry ingestion
GrafanaFleet operational dashboards
Microsoft FabricReal-Time Intelligence KQL analytics

🏗️ Architecture

LayerComponentDescription
EdgeTelemetry AgentCollects inference metrics and health data on each robot
TransportIoT OperationsMQTT broker and edge-to-cloud routing
IngestionEvent HubsCloud endpoint for partitioned telemetry streams
AnalyticsFabric RTIKQL queries for fleet-wide trend analysis
VisualizationGrafanaReal-time dashboards and alert rules
AutomationDrift DetectionStatistical monitoring for policy degradation
AutomationRetraining TriggersAutomated training pipeline initiation
GuideDescription
Telemetry SpecificationSchema and routing architecture
Dashboard SpecificationFleet dashboard and alerting design
Drift Detection SpecificationDetection algorithms and thresholds
Retraining SpecificationAutomated retraining trigger pipeline