Appearance
Validation
Every merge runs the standard manoeuvres for each reference vehicle and compares them against golden results with tolerances. Physics-visible changes must update the golden file in the same pull request and say so in a physics: changeset line.
Current results
Core 0.8.0 · math self-test hash ed4e6b369d1eeace
| Vehicle | Kus four-wheel (deg/g) | Kus single-track (deg/g) | Kus linear theory (deg/g) | 0–100 km/h (s) | 100–0 km/h (m), locked / ABS | Mean decel (m/s²) | Wheels lock | Parks on slopes | Timestep sweep | Step steer: yaw-rate response (s) / overshoot (%) | Lane change passes up to (km/h) | 100–0 km/h on ice (m), locked / ABS | Scripted drive hash |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| hatchbackFwd | 1.08 | 0.93 | 0.93 | 9.82 | 46.5 / 42.6 | 8.22 | at 0.33 s , no chatter, ripple 0.0 %, at rest after the stop (5e-5 m/s) | yes (worst creep 9.3e-7 m/s) | stable (Kus ±0.001 deg/g, braking ±0.3 %) | 0.21 / 4.3 | 100 | 331 / 330 | b818efc68f41046d |
| sportsRwd | 0.71 | 0.59 | 0.58 | 4.94 | 38.7 / 35.5 | 9.96 | at 0.20 s , no chatter, ripple 0.0 %, at rest after the stop (8e-6 m/s) | yes (worst creep 9.6e-7 m/s) | stable (Kus ±0.001 deg/g, braking ±0.0 %) | 0.20 / 0.5 | 110 | 282 / 281 | 35c85bdc3695341a |
| kart | -0.21 | 0.31 | 0.36 | 10.87 | 58.0 / 54.8 | 6.34 | at 0.79 s , no chatter, ripple 0.0 %, at rest after the stop (2e-11 m/s) | yes (worst creep 1.1e-16 m/s) | stable (Kus ±0.000 deg/g, braking ±0.6 %) | 0.15 / 1.9 | 80 | spins (225 / 226) | f8d2f902788bf586 |
| pickup4x4 | 0.94 | 0.72 | 0.71 | 7.44 | 52.1 / 47.1 | 7.37 | at 0.26 s , no chatter, ripple 0.0 %, at rest after the stop (2e-5 m/s) | yes (worst creep 1.1e-9 m/s) | stable (Kus ±0.001 deg/g, braking ±0.1 %) | 0.22 / 2.5 | 90 | 361 / 360 | 36ed6dc9c3039dc1 |
| crossoverEv | 0.29 | 0.22 | 0.21 | 5.03 | 44.5 / 40.1 | 8.65 | at 0.38 s , no chatter, ripple 0.0 %, at rest after the stop (8e-6 m/s) | yes (worst creep 2.5e-7 m/s) | stable (Kus ±0.000 deg/g, braking ±0.3 %) | 0.25 / 0.3 | 90 | 330 / 330 | 90eea945a7518680 |
| openWheeler | 0.64 | 0.60 | 0.66 | 2.90 | 24.0 / 22.5 | 15.43 | at 0.06 s , no chatter, ripple 0.0 %, at rest after the stop (1e-14 m/s) | yes (worst creep 3.8e-8 m/s) | stable (Kus ±0.000 deg/g, braking ±0.3 %) | 0.14 / 0.0 | 100 | 166 / 165 | d7cc5adbd593f5a0 |
Understeer gradient follows ISO 4138 (constant radius, 40 m, speeds 4 to 12 m/s), run on the four-wheel model and, as a cross-check, on the single-track model from the same definition. Each point's steer angle is reduced by the Ackermann angle of the path actually driven, L · r / V, before the fit, so the small speed errors a drivetrain leaves at part throttle do not bias the slope. The "linear theory" column is K_us = (m / L') · (b' / C_f − a' / C_r) with axle cornering stiffnesses at their static loads and moment arms corrected for pneumatic trail. The single-track model agrees with it within a few hundredths of a degree per g; the residual is tire nonlinearity across the measured lateral-acceleration range. The four-wheel model sits a few tenths higher because lateral load transfer, split by roll stiffness, costs the more heavily loaded axle grip through load sensitivity (load transfer), and the sports car's limited-slip differential adds a little more under power. The kart is the exception: its solid rear axle (drivetrain) forces both rear wheels to one speed, so in a corner the inner wheel drives and the outer brakes, a yaw moment against the turn that doubles the steer angle needed at low speed. The push eases as load transfer and steering jacking (ADR-0012) unload the inner rear wheel, so a linear fit over lateral acceleration reads it as a negative gradient: the number is a poor summary for a solid axle, and the per-point steer angles in the golden file tell the story (twice the Ackermann angle at 4 m/s, falling with speed). The single-track model, with one wheel per axle, shows the kart's underlying understeer.
Straight line reports 0–100 km/h at full throttle, through the preset's drivetrain (the automatic launches on its clutch and shifts up, so the time includes wheelspin and the torque holes), and 100–0 km/h at full brake with no ABS, so cars whose brakes exceed tire grip lock their wheels and stop on sliding friction. Published road-test distances assume ABS; the table also shows the same stop with the ABS assist on (assists), which brings the sports car to within a few metres of its data sheet.
Locked brakes is measured on that same stop. A wheel that locks must lock once and stay locked: the scenario counts every change between rolling and locked at substep resolution and reports the releases while the car is still moving, which is lock chatter; the reference vehicles show none. The driven wheels lock last, because they also have to drag the engine's reflected inertia down through the clutch; the kart's single-speed engine from near redline takes the longest. The deceleration on sliding friction is reported as its relative RMS ripple about a 0.2 s moving average (below 0.01 % for every preset; the slow drift with speed from aero drag and the friction curve is not counted). After the stop the brake stays held for two seconds: the sliding tires release the contact-patch deflection they stored (ADR-0010), a spring-back of a few centimetres at under 0.3 m/s, and the car must then be at rest below 0.1 mm/s.
Parks on slopes runs the standstill scenarios: at rest on flat ground with no inputs, parked facing uphill on 10 %, 20 % and 30 % grades on the service brake, on 10 % and 20 % grades on the handbrake alone (skipped for vehicles without one), and across a 20 % slope with both held. After a 5 s settle the car must sit below 0.1 mm/s, drift less than 1 mm over the next 10 s and show no sustained oscillation (velocity RMS below 0.1 mm/s over the last 5 s). The table shows the worst creep speed over the cases.
Timestep sweep runs the skidpad (three speeds), the locked-wheel stop and a parked hold on a 30 % grade at every combination of internal substep rate (250, 500, 1000, 2000 Hz) and host step rate (30, 60, 120, 240 Hz), and compares each cell with a reference cell at the definition's own substep rate and the 100 Hz host step the other scenarios use. A vehicle passes when every cell is finite, parks, stops without chatter and comes to rest, the understeer gradient stays within 0.05 deg/g of the reference and the braking distance within 1 %. The presets sit at a thousandth of a degree per g and under 1 %. On the built-in host the host rate only changes how often the inputs update (the proxy integrates at the substep rate), so the sweep is mostly a check on the substep rate; the host-rate dimension is there for the external-host contract, where the impulse exchange runs at the host rate.
Step steer follows ISO 7401: the car runs straight at 80 km/h, then the road-wheel angle is ramped over 0.1 s (the standard allows up to 0.15 s; a ramp keeps the result independent of which host step the input lands on) to the angle linear theory needs for 4 m/s² of lateral acceleration, with the definition's own cornering stiffnesses and trail, and held for 5 s. The table shows the time for the yaw rate to first reach 90 % of its steady state and its overshoot past it; the golden file also has the steady yaw rate and its gain per radian of steer, the lateral-acceleration response time, the steady side-slip angle and the roll. A short wheelbase and a small yaw inertia answer fastest: the kart in under two tenths of a second, the road cars in two to three. The lateral acceleration comes out a little under the 4 m/s² aimed for because the steer angle is the linear one and the tires are not.
Double lane change lays out the ISO 3888-1 course: 15, 30, 25, 25 and 15 m sections with lane widths of 1.1, 1.2 and 1.3 times the vehicle width plus 0.25 m and a 3.5 m offset between the lane centres (the ISO 3888-2 "moose test" course is an option). The vehicle width defaults to the track width plus 0.25 m. A scripted driver steers through at each entry speed from 50 to 110 km/h in steps of 10: path-curvature feedforward with the linear understeer gradient, pure pursuit on a smooth centreline, and up to four practice runs per speed in which it learns a steering correction along the course from the path error of the previous run, as a test driver does; the best run counts. A speed passes when every wheel stayed inside the coned lanes of sections 1, 3 and 5, and the table shows the highest speed that passed. The number depends on this driver as much as on the car, so compare presets against each other and against road tests only loosely; the per-speed attempts with their cone overlap, peak lateral acceleration, yaw rate and side slip are in the golden file.
Surfaces repeats the 100–0 km/h stop on wet asphalt, gravel, snow and ice from the reference table, with locked wheels and with the ABS. The distance scales roughly with the inverse of the grip, so ice at 0.12 takes about seven times the dry distance; the ABS gains less on ice than on asphalt because the surface scales the friction and not the stiffness, so the force peaks at a smaller slip and the ABS's default slip target of 0.12 sits past it. The table shows the ice column. A car with rear-only brakes locks its rear wheels and swaps ends on snow and ice (the kart does); the scenario flags it as spun, the distance is then what it travelled before coming to rest, and the table shows "spins" instead of a number.
Scripted drive hash is the state hash after a fixed 30 s drive. It changes whenever anything physics-visible changes, and it must match across Chromium, Firefox, WebKit, and Node. The determinism harness also replays a recorded lap of the sandbox track for each preset; see the determinism contract.
Against an independent simulator
The scenarios above check Skidpad against linear theory, against itself across timesteps, and against published road-test figures the presets were tuned to. An independent check drives Skidpad and Project Chrono's multibody BMW E90 through the same manoeuvres with the same inputs, with the Skidpad car built from Chrono's published constants. Steady-state handling agrees within a few percent: understeer gradient, lateral acceleration to the limit, yaw gain, roll and turn radius. The comparison found Skidpad answering a step steer too quickly because it could not express the car's static toe, and pitching too little because it had no anti-dive geometry. With staticToeDeg, antiBrake and antiDrive derived from the hardpoints, and the car's own engine-braking map, the step and sine responses agree too and the braking pitch gap halves. The remaining differences are geometry that changes with travel. The report has the full results, and tools/chrono-compare reruns them.
A second Chrono car, a front-wheel-drive sedan with no anti-roll bars and a Magic Formula tire, is built with nothing fitted to its behaviour, so its roll gradient and load transfer are predictions. It agrees on 22 of 27 metrics. Where both cars miss in the same way (about 0.2°/g less understeer, braking pitch that settles too quickly, a front that toe-in forces lift further than in Chrono), the cause lies in the engine or in Chrono rather than in either car. Its report compares the two.
Against a measured car
The first comparison against a real car uses NHTSA's instrumented 1997 Jeep Cherokee, whose slowly increasing steer, step steer and two lane changes are read exactly from the vector plots of SAE 2000-01-0700. The car's measured parameter set is not public, so it is built from published specifications and estimates, and three unknowns are fitted on the slowly increasing steer alone: the front anti-roll bar, the tire's peak friction and the steering's compliance steer under the kingpin torque (alignTorqueComplianceDeg, ADR-0027), which this comparison showed Skidpad was missing. With them the step steer and both lane changes, at 12 and 22.5 m/s, are predicted on every metric and trace within tolerances set before the first run. The report records the estimated build's misses and every change made after the first run, and tools/nhtsa-compare reruns it.
Coming with later milestones
A rest-jitter measurement on an external host: the standstill scenarios run on the built-in host, and the external-host contract has only the timestep sweep's host-rate dimension to cover it.