Fleet Autonomy Architecture

vehicle · depot edge · cloud · ODD

A fleet's autonomy is a distributed system, not a feature in each vehicle. The depot is the tier that ties the vehicle to the training cluster.

Fleet autonomy runs on three compute tiers. The vehicle runs perception through control in milliseconds; the depot edge handles log ingest and OTA staging over minutes to hours; the cloud runs training and simulation over days to weeks. Logs flow from vehicle to depot to cloud, and models, maps and policies flow back. Below, the on-vehicle stack: sense, perceive, predict, plan and control, with localization alongside; end-to-end models merge the middle stages. FLEET AUTONOMY RUNS ON THREE COMPUTE TIERS Vehicle perception to control milliseconds Depot edge log ingest, OTA staging minutes to hours Cloud training, simulation days to weeks logs flow right; models, maps and policies flow back left ON-VEHICLE STACK Sense Perceive Predict Plan Control Localization runs alongside; end-to-end models merge the middle stages.
Compute tiers
3
Vehicle loop
Milliseconds
Training loop
Days to weeks
Hub
Autonomy

Finding

For a fleet, autonomy is a three-tier system - vehicle, depot and cloud - and the operator's real design choices sit in the depot and the operating domain, not in the driving software

Every robotaxi and driverless truck runs the same on-board chain - sensing, perception, prediction, planning and control - on an inference computer such as Nvidia DRIVE Thor, Tesla's AI4 and AI5 or Horizon Robotics' Journey chips. Tesla and the Chinese leaders have collapsed the middle of that chain into end-to-end neural networks, which shifts the hard problem to training data and compute. The fleet adds two tiers the consumer car does not need: depot edge systems that move logs off the vehicle during dwell, and cloud clusters that train, simulate and validate each release.

Fleet autonomy differs from consumer autonomy because the operator controls the operating design domain (ODD) - the roads, speeds, weather and hours the system is approved for. Waymo, Baidu Apollo Go and Aurora each launched in a bounded ODD and widened it city by city or lane by lane. Remote assistance staff resolve situations the vehicle cannot, and fallback rules - pull over, stop in lane, return to depot - are part of normal operations rather than exceptions.

The architecture sets operating cost. Sensor suite choices fix calibration and spare-parts work; Waymo cut its camera count from 29 to 13 between its fifth- and sixth-generation drivers. Depot connectivity and edge storage set how much data can leave each vehicle per shift. ODD boundaries set how many remote assistants a fleet needs. See Edge Compute and Data Pipeline and Depot and Yard Integration.

Stack layers

On-vehicle stack layers

Fleet autonomy stack layers
LayerRoleFleet consideration
SensingCameras, radar, lidar, IMU, wheel odometryOne sensor suite per fleet simplifies calibration and spares
PerceptionObjects, lanes, drivable spaceTuned to the ODD: urban core, yard or highway
PredictionWhere other road users go in the next secondsDepot, dock and bus-stop behaviors matter
PlanningRoute, behavior and trajectoryTied to dispatch, depot location and charger availability
LocalizationPosition from GNSS, IMU and map matchingPrecise docking at chargers and bays
ControlDrive-by-wire steering, braking, regenerative brakingRide comfort, cargo stability, wear and energy use

Compute tiers

Where each workload runs

Fleet autonomy compute tiers
TierWorkloadsConstraintExamples
VehicleReal-time inference for the driving stackPower, heat, latency, functional safetyNvidia DRIVE Thor, Tesla AI4/AI5, Horizon Journey, Qualcomm Snapdragon Ride
Depot edgeLog ingest, filtering, OTA staging, calibration, diagnosticsDwell time, network capacity, storageServers and storage at each depot
CloudTraining, simulation, regression tests, safety-case evidenceTraining compute and powerTesla Cortex (Austin), Waymo on Google TPUs, Baidu and XPeng clusters

Operations

ODD, fallback and remote assistance

Fleet autonomy operating controls
ControlWhat it definesOperator lever
Operational design domainRoads, speeds, weather, hoursStart narrow; expand by evidence
Fallback behaviorMinimal-risk maneuvers when conditions driftPull over, stop or return to depot
Remote assistanceHuman guidance for edge casesStaffing ratio scales with ODD complexity
Health monitoringSensor, compute and actuator statusPull vehicles before faults reach the road
Safety caseEvidence for regulators and insurersLogged data and simulation results

Sources

Sources and scope

Scope
Fleet Autonomy Architecture
Sources
Waymo (sixth-generation driver); Nvidia; Tesla; Horizon Robotics; Baidu; Aurora; ElectronsX framework
Updated
October 2026