Skip to content

Trajectory Tracking Control

Purpose: Map the boundary between planning and control for autonomous vehicles, covering nominal path/trajectory tracking, vehicle dynamics models, controller families, safety filters, and deployment tests for indoor, outdoor, and airside operations.

Key Takeaway: A planner is only deployable if its trajectory can be tracked under actuator delay, saturation, tire slip, localization noise, payload variation, degraded surfaces, and emergency fallback. Airside vehicles should treat tracking control as a safety-critical subsystem with its own metrics, fault injection, and Simplex-compatible fallback, not as a generic PID afterthought.

Research current as of: 2026-05-09


Problem Framing

Planning answers "where and when should the vehicle go?" Control answers "what steering, throttle, brake, and gear commands make the vehicle actually do that?" The interface looks simple: a trajectory of poses, velocities, accelerations, and timestamps. In deployment, this is where most physical assumptions become visible.

For airside vehicles, the control problem is dominated by low-speed precision, frequent stop-and-go operation, tight turning near aircraft, heavy or changing payloads, hydraulic/electric actuator delay, wet apron surfaces, GNSS/localization covariance, and the need for predictable comfortable motion around workers. Tracking error is not just a comfort issue: a 0.5 m lateral miss can be acceptable on a wide road and unacceptable near a cargo door, belt loader, or aircraft safety envelope.

The controller should therefore be evaluated as part of the autonomy stack:

  • Can it track nominal trajectories within clearance margins?
  • Does it expose when a planned trajectory is not physically trackable?
  • Does it degrade safely under delay, saturation, model mismatch, low friction, or actuator faults?
  • Does the safety monitor know when to slow, replan, switch controllers, or stop?

Method and Architecture Taxonomy

Planner-Control Interface

text
Planner output
  -> time-indexed trajectory: pose, heading, velocity, acceleration, curvature
  -> semantic tags: stop point, yield point, docking phase, reverse phase, no-go margin
  -> feasibility envelope: max curvature, speed, acceleration, jerk, tracking margin

Controller input
  -> trajectory
  -> current state estimate and covariance
  -> vehicle dynamics parameters
  -> actuator state and health
  -> surface / payload / ODD mode

Controller output
  -> steering, throttle/brake, gear, optional control horizon
  -> predicted tracking rollout
  -> health and confidence flags

Autoware's trajectory follower pattern is a useful production reference: lateral control can be MPC or pure pursuit, longitudinal control is typically PID, and debug topics expose predicted trajectories and errors. Airside deployments should preserve this observability and add domain-specific stop, docking, and low-speed reverse metrics.

Vehicle Models

ModelUseStrengthLimit
Kinematic bicycleLow-speed path tracking and planning feasibilitySimple, interpretable, fastIgnores tire slip, load transfer, actuator dynamics
Dynamic bicycleHigher speed, low-friction, payload-sensitive controlCaptures lateral dynamics and slipNeeds tire/cornering parameters and careful identification
Actuator-delay modelSteering/brake/throttle lag compensationDirectly addresses deployment mismatchMust be measured per platform and maintained
Learned residual modelCorrects systematic model errorUseful for buses, tugs, heavy payloads, worn actuatorsNeeds bounded use and fallback
Full multibody / tire modelOffline simulation, edge cases, certification evidenceHigh fidelityToo expensive and parameter-sensitive for normal control loop

Controller Families

ControllerBest FitDeployment Notes
PID / cruise controlLongitudinal speed and stop-point trackingNeeds feedforward, anti-windup, saturation handling, and grade/load compensation
Pure pursuitSimple low-speed lateral trackingStable and easy to tune; can cut corners and oscillate if lookahead is wrong
StanleyCross-track plus heading correctionGood for lane/path tracking; gain scheduling needed at very low speed
LQRLinearized tracking around a referenceTransparent and fast; sensitive to model validity
MPC / NMPCConstraint-aware tracking with actuator limitsBest general-purpose controller when compute and model quality are sufficient
MPPI / sampling MPCNonlinear and nonconvex costs, multiple constraintsUseful for difficult dynamics; sampling count and GPU budget matter
iLQR / smart MPCFast trajectory optimization with learned residualsAttractive for vehicles whose true dynamics differ from nominal
CBF / safety filterLast-mile invariant safety constraint enforcementShould sit around or below nominal control, not replace planning
Simplex trackingHigh-performance controller plus verified baselineLets learned/adaptive control run while a high-assurance fallback remains available

Nominal Airside Stack

text
Behavior / motion planner
  -> trajectory validator
  -> MPC or pure-pursuit + PID nominal controller
  -> CBF / collision and envelope safety filter
  -> command gate / DBW interface
  -> actuator health monitor
  -> controller evaluator
  -> planner feedback: trackable, degraded, stop, replan

A practical starting point for low-speed airside vehicles is MPC lateral control plus PID longitudinal control, with pure pursuit as a simple fallback and CBF/Simplex monitoring for safety-critical envelopes. Smart MPC or learned residuals are useful after collecting enough platform-specific data to identify actuator delay, payload effects, and surface-dependent slip.


Evaluation and Deployment Notes

Metrics

CategoryMetrics
Accuracylateral error, heading error, speed error, stop-point error, docking final pose error
Stabilityoscillation, overshoot, settling time, sign changes in steering, recovery after disturbance
Comfortlongitudinal/lateral acceleration, jerk, yaw rate, steering-rate limits
Feasibilitycurvature limit violations, acceleration/brake saturation, reverse/gear transition correctness
Safetyminimum clearance after tracking error, CBF interventions, emergency stop distance, watchdog trips
Robustnessscore under delay, low friction, payload change, localization covariance, actuator derate
Operationscontroller faults per hour, degraded-mode time, manual takeover / remote assistance triggers

Test Matrix

  1. SIL regression: replay planned trajectories through a vehicle model with injected delay/noise.
  2. Controller-in-the-loop: run the real control node against simulated odometry and actuator dynamics.
  3. HIL/DBW bench: verify command timing, saturation, watchdogs, command gate, and fault flags.
  4. Closed-course tests: straight, curve, S-curve, stop line, docking, reverse, wet surface, payload variants.
  5. Airside supervised tests: stand approach, service-road turns, hold-short stops, interaction with human-driven GSE.

For each release, the planner should be checked against the controller envelope. A trajectory that requires curvature, acceleration, jerk, or stop accuracy outside the controller's validated envelope should be rejected before actuation.

Deployment Practices

  • Identify steering, brake, and throttle delay from bag data and re-check after maintenance.
  • Use timestamp provenance and controller deadlines; stale trajectory or odometry should trigger controlled stop.
  • Keep lateral and longitudinal controllers synchronized; mismatched command ages can produce unsafe combined motion.
  • Log planned trajectory, executed trajectory, actuator commands, saturations, and controller state on every run.
  • Separate "planner unsafe" from "controller unable to track" in incident root-cause labels.
  • Tune by ODD mode: indoor dock, outdoor yard, apron transit, stand approach, reverse docking, degraded weather.
  • Add command-gate limits for max speed, steering rate, acceleration, deceleration, and jerk by zone.

Indoor / Outdoor / Airside Fit

DomainFitController Emphasis
Indoor AMR / forkliftVery highLow-speed precision, pallet/load stability, tight aisle turns, certified safety scanner stops
Outdoor industrial yardVery highTrailer/payload variation, uneven surfaces, GNSS covariance, rain/snow traction
Public-road AVHighHigher-speed lateral stability, lane keeping, traffic speed control, comfort
Airside apronVery highStand clearance, repeated stop-go, aircraft proximity, wet apron, ground crew comfort
Movement area / runway-adjacentHigh with strict gatesStop-line accuracy, authority-dependent speed caps, default safe stop on command uncertainty

Airside control is forgiving in speed but unforgiving in geometry. Low speed creates time for validation and fallback, but the clearance envelope near aircraft, equipment, and personnel makes tracking-error budgets smaller than they appear from road-AV intuition.


Failure Modes

Failure ModeSymptomMitigation
Actuator delay not modeledOvershoot at stop points, corner cutting, late brakingDelay identification, feedforward, MPC delay compensation
Saturation ignoredController cannot achieve planned curvature or decelerationPlanner-controller envelope check and command limiting
Localization covariance spikeController chases noisy pose or diverges near buildingsCovariance-aware gains, speed reduction, fallback localization policy
Low-friction apronABS/traction events, lateral slip, longer stopping distanceSurface mode, friction estimation, inflated braking margins
Payload / tow changeUndersteer, slow acceleration, poor stop accuracyPayload-aware parameters, adaptive/residual model with bounds
Reverse or docking mode bugsWrong sign on steering, poor final alignmentDedicated low-speed reverse controller and mode-specific tests
Controller fights safety filterOscillation between nominal command and safety overrideExplicit priority architecture and post-filter predicted rollout
Planner emits untrackable trajectoryHigh controller effort, repeated replans, emergency stopFeasibility validation before trajectory acceptance
Stale input dataCommands based on old trajectory or odometryTimestamp checks, timeout thresholds, controlled stop
Learned controller out of distributionSmooth behavior in training set, unsafe in rare casesSimplex architecture, ODD monitor, high-assurance fallback


Sources

Public research notes collected from public sources.