Quix Webinar · Eng. Problem · DWG-A / SCHEMATIC BLUEPRINT
SHEET 01 / 08 · THE PROBLEM

The Engineering Problem sim-to-real correlation.

REV A · 2026-06-04

You tune against a model. Then reality disagrees. Closing that gap — and knowing where the model can't be trusted — is the job.

SHEET 02 / 08 · DATA SOURCE

The data — real car data from CAN bus.

SOURCE · COMMA.AI OPENPILOT

Comma.ai openpilot logs. This is what the car's own controllers saw: the same signals openpilot drives on. Production hardware, real roads.

└─data/raw/segments/
  └─PLATFORM/  # Ford, Hyundai, Tesla, …
    └─device/  # comma 3 dongle ID
      └─route/  # one drive
        └─segment/  # 60 s window
          └─rlog.zst  # raw CAN log

One segment ≈ 60 seconds of driving · 3,000 samples per channel · steering, speed, yaw rate, lateral accel — sampled at 50 Hz.

SHEET 03 / 08 · VEHICLE INVENTORY

Four cars — one model is about to treat all of these the same.

N = 4 PLATFORMS
PLAT-01 · TESLA

Tesla Model 3

wheelbase 2.875 m
mass 2,035 kg
PLAT-02 · FORD

Mustang Mach-E

wheelbase 2.984 m
mass 2,336 kg
PLAT-03 · HYUNDAI

Ioniq 5

wheelbase 2.970 m
mass 2,084 kg
PLAT-04 · FORD

F-150 Lightning

wheelbase 3.70 m
mass 3,084 kg

A 3-ton truck and a compact hatch. The baseline model is mass-independent — it predicts identical cornering for both.

SHEET 04 / 08 · VIRTUAL MODEL

The virtual model — KS, the "driving-school" model.

KINEMATIC SINGLE-TRACK

A rigid rod of wheelbase L, no tyre, no slip — the car goes exactly where the wheels point. Yaw rate falls straight out of geometry.

ψ̇ = (v / L) · tan(δ)
a_y = v · ψ̇ // no forces computed
STATE x = (x, y, ψ, v, δ)
INPUT u = (δ̇, a)
SHEET 05 / 08 · SCOPE OF MODEL

We give it speed and steering. It returns the lateral response.

CLAMPED · LATERAL ONLY
INPUT · MEASURED

What goes in

v — longitudinal speedm/s · clamped
δ — road-wheel anglerad · clamped
KS MODEL
PREDICTED · LATERAL

What comes out

x — positionm
y — positionm
ψ — headingrad

Measured v and δ are clamped at every integration step, so the longitudinal channel is an input, not a prediction. What's left is purely lateral — and the model's lies are all lateral, so this isolates exactly the residual we want to measure.

SHEET 06 / 08 · V0 · sim.csv SCHEMA

How V0 is built · the file — one row per 50 Hz sample, four kinds of column.

50 Hz · WIDE TABLE
01 · INPUTS
delta_road_rad
v_mps
a_long_mps2
accel_pedal_pct
From CAN — given to the model at inference.
02 · TRUTH
yaw_rate_meas_rads
a_lat_meas_mps2
·
·
Ford / Hyundai only. Tesla doesn't expose these.
03 · PREDICTION
yaw_rate_pred_rads
a_y_pred_mps2
x_m
y_m
psi_rad
04 · RESIDUAL
yaw_rate_resid_rads
a_y_resid_mps2
·
·
= pred − truth. The thing we want to shrink.

The truth columns exist in sim.csv — but not in what your model is given at inference.

SHEET 07 / 08 · THE TASK

The task — predict the lateral response better than V0.

TWO METRICS · SCORED
METRIC · 01 · INSTANTANEOUS

Yaw-rate RMSE (rad/s)

Instantaneous fidelity — how close the predicted yaw rate is to measured, every sample.

METRIC · 02 · INTEGRATED

Cross-track-error RMSE (metres)

Where the integrated trajectory actually ends up, resampled at uniform distance.

Not redundant: a tiny persistent yaw bias is nearly invisible per-sample but compounds into hundreds of metres of drift.

SHEET 08 / 08 · BASELINE V0

The start line — beat these two numbers.

N = 534 SEGMENTS · HELD-OUT

V0, scored on 534 held-out segments. Everything from here is measured against them.

YAW-RATE RMSE
0.0163
rad / s
CROSS-TRACK-ERROR RMSE
254
metres

The 254 m is the compounding-bias problem made concrete — integrate a slightly biased yaw rate over a minute of driving and the trajectory drifts off the map.

← → · click halves · space