Compare path-tracking controllers. Inspect the motion, understand the error, and test your assumptions.
Equal-axis scale · world coordinates in meters
Local normal error · left of the route is positive
Shared replay cursor · individual completion times
All four views use the same spatial scale and replay time. Use the Live workspace timeline to inspect a particular moment.
Four lateral controllers share one reference, one initial condition, one speed policy and one constrained plant. Every value displayed is computed from this experiment. These are illustrative kinematic simulations, not physical robot measurements or evidence that a controller is universally superior.
Tracks the bearing to a forward target. Integral action is limited during saturation and the derivative is filtered to reduce abrupt commands.
A target-bearing controller: lookahead influences corner cutting, while gain and actuator lag influence oscillation.
Uses the circular arc from the robot to a forward target. The denominator is the actual robot-to-target distance, rather than an assumed fixed distance.
Long lookahead tends to smooth the response and cut corners; short lookahead can increase steering demand.
Computes feedback gains from the discrete Riccati equation for a linearized lateral and heading error model. Curvature feedforward supplies the nominal turn command.
The linearization assumes small errors. Physical clipping and actuator lag are outside this two-state design model; constrained optimality is not implied.
Rolls out finite control sequences through the constrained motion model, evaluates a horizon cost, then applies only the first command and replans.
Here u is the requested yaw rate and Δu its change. A bounded, warm-started candidate search approximates the solution. It does not certify a global optimum, recursive feasibility, or closed-loop stability.
Body dimensions define the displayed footprint only. No collision geometry, wheel torque, tire slip, sensor uncertainty, localization estimator, terrain contact, or implement dynamics is simulated. Differential drive uses a body-level unicycle model; individual wheel-speed limits are not modeled.
Route feasibility: raw corners have discontinuous heading and are intentionally a stress test. Smoothed routes can still exceed a robot’s curvature limit. Slowing an Ackermann robot cannot overcome a geometric turning-radius constraint.
Plot versus table: the chart and inspector show the signed local normal component; the table uses unsigned shortest distance to any reference segment. These can differ at endpoints and path crossings.
RMSE / P95 / peak: shortest Euclidean distance to any segment of the reference polyline over the recorded run; lower errors only have meaning alongside route progress and completion. Errors from an unfinished run are not full-route performance.
Completion: reported only when terminal position and route-progress criteria are met and achieved speed is at most 0.12 m/s. A timeout is displayed explicitly; it is never assigned a successful completion time. Saturation: fraction of steps activating either the turn-rate envelope or the yaw-rate slew limit. Yaw effort: ∫ω²dt in rad²/s; a control activity indicator, not energy consumption.
Fair comparison: common plant and scenario do not imply equally optimal tuning. Different controllers use different preview models. Compare several routes, initial errors and parameter choices before drawing conclusions; no statistical significance is claimed from one deterministic run.
These are the effective settings of the displayed run. The JSON export contains every numerical setting and all original waypoints.
These references support the underlying methods. They do not validate this simulator’s implementation or its default tuning.
Final metrics from the complete simulation · independent of the replay cursor
| Controller | Outcome | RMSE (m) | P95 (m) | Peak (m) | Time (s) | Progress | Goal error (m) | Saturated |
|---|---|---|---|---|---|---|---|---|
| Run an experiment to generate results. | ||||||||