Skip to content

Development & delivery

Digital twins and simulation

What a digital twin is once the marketing is removed, the four kinds that actually get built, and how to tell a useful one from an expensive demonstration.

Intermediate8 minDigital twin · Simulation · Testing

"Digital twin" is among the most oversold phrases in the industry. There is a genuinely useful idea underneath, and separating it from the demo footage is worth doing carefully.

The definition, stripped down#

The four kinds that actually get built#

Programmes say "digital twin" and mean quite different things. Naming them prevents a lot of talking past each other.

1. Component twin#

A model of one part — a battery pack, a motor, a brake actuator.

2. Vehicle twin#

A model of one complete vehicle — its configuration, software versions, faults and history.

3. Environment twin#

A simulated world — roads, other traffic, weather, sensor physics.

4. Fleet twin#

An aggregate model across many vehicles.

Gotcha

When someone proposes a digital twin programme, ask which of the four they mean and what decision it will change. "All of them" usually means none of them well.

The kind with the fastest payback is nearly always the vehicle twin — because it is data engineering rather than physics, and because it immediately improves OTA targeting and diagnostics.

What a twin is worth in cockpit work#

You are unlikely to build the physics. You will still benefit in four concrete ways.

Realistic test data. A battery twin driving your SIL environment produces plausible state-of-charge curves — including the awkward non-linear end of discharge where display bugs hide. Far better than a linear ramp.

Diagnosing a customer complaint. "The range estimate jumped by 40 km" is untestable as a bug report. With a vehicle twin holding that car's software versions, configuration and recent signal history, it becomes reproducible on a desk.

Deciding who gets an update. A vehicle twin knows which hardware revision and which dependent components each car has, which is what makes a targeted rollout possible rather than a fleet-wide gamble.

Answering questions about the past. "Did this ever happen before the March release?" is a query against fleet history, not a testing exercise.

How faithful does it need to be#

The most common failure is building for more fidelity than the decision requires.

Question being askedFidelity needed
Does the UI render correctly at 5% charge?Very low — set a number
Does the charging state machine handle interruption?Low — a state model
Does the range estimate behave in winter?Medium — thermal battery model
Will the pack meet warranty in 8 years?High — validated ageing model
Will the perception stack detect a child at dusk?Very high, and still not sufficient alone

How to tell a real twin from a demonstration#

Questions that separate the two quickly:

"Where does its state come from, and how often?" A real twin has a telemetry path and an update rate. A demonstration has a configuration file.

"How was it validated?" Against what recorded reality, and what was the error? An unvalidated model is a hypothesis.

"What decision does it change?" If nobody acts differently because of it, it is a visualisation.

"What happens when the vehicle is offline for a month?" Real fleets go quiet. A twin that assumes continuous connectivity has not met one.

"Can I query one VIN?" If the answer is only aggregate dashboards, it is analytics, which is fine — but it is not a vehicle twin.

Gotcha

A rendered 3D vehicle rotating on a screen is a visualisation. It may sit on top of a real twin, or on nothing at all. The graphics tell you nothing about whether the underlying state is synchronised, validated, or used.

This sounds cynical. It is the single most useful filter to apply in this area.

Starting small, in the direction that pays#

If you want the benefit without a programme-scale investment:

  1. Record real traces and replay them. This is a twin of one drive, it is perfectly faithful by construction, and it costs nothing.
  2. Version the vehicle configuration — software versions, hardware revisions, options — in a queryable store keyed by VIN. This alone unlocks targeted OTA and most diagnostics value.
  3. Add a physics model only where a decision needs one, and validate it against recorded data before trusting it.
  4. Keep signal history for the fleet, at a sample rate you can afford. Most "we cannot reproduce it" tickets die here.

Next#

Getting software onto vehicles that are already sold — OTA in an SDV.

References & further reading

Code links target the main branch on cs.android.com. AOSP moves — if a path 404s, search the symbol instead.