Quix Webinar · The Engineering Problem
The Engineering Problem

Sim-to-real
correlation.

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

GAP
model · KS reality · measured residual
The Data

Real car data
from the CAN bus.

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

No simulator. No synthetic disturbances. Every yaw rate, every steering angle came off the vehicle bus at 50 Hz.

data/raw/segments/ PLATFORM/ device/ route/ segment/ rlog.zst
Four Cars

One model is about to
treat all of these the same.

Tesla Model 3
wheelbase 2.875 m · sedan
2,035 kg
Ford Mustang Mach-E
wheelbase 2.984 m · crossover
2,336 kg
Hyundai Ioniq 5
wheelbase 2.970 m · crossover
2,084 kg
Ford F-150 Lightning
wheelbase 3.70 m · full-size pickup
3,084 kg

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

The Virtual Model

KS — the
driving-school model.

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
L · wheelbase v δ ψ
The Scope

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

Input · measured

What we feed in

  • v  — speed, clamped
  • δ  — road-wheel angle, clamped
Predicted · lateral

What it returns

  • x  — position
  • y  — position
  • ψ  — heading

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.

How V0 is Built · The File

One row per 50 Hz sample,
four kinds of column.

Inputs
  • delta_road_rad
  • v_mps
  • a_long_mps2
  • accel_pedal_pct
Truth
  • yaw_rate_meas_rads
  • a_lat_meas_mps2
Ford / Hyundai only
Prediction
  • yaw_rate_pred_rads
  • a_y_pred_mps2
  • x_m
  • y_m
  • psi_rad
Residual
  • yaw_rate_resid_rads
  • a_y_resid_mps2
= pred − truth

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

The Task

Predict the lateral response
better than V0.

Metric 01
Yaw-rate RMSE (rad/s)
Instantaneous fidelity — how close the predicted yaw rate is to measured, every sample.
Metric 02
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.

The Start Line

Beat these two numbers.

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

Yaw-rate RMSE
0.0000
rad / second
Cross-track-error RMSE
0
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.

arrow keys · click halves