A Low-Cost, Open Platform for End-to-End Autonomous Driving on a Miniature Ackermann Vehicle

The 30-second take

• What: The authors release a low-cost platform and a command-conditioned behavior-cloning baseline that reaches 6.1 cm mean cross-track error on the real track (human demo 4.7 cm), and they show FOV and sim-to-real data choices decide whether all routes complete.
• Why it matters (abundance angle): End-to-end driving research is still capital-heavy. A cheap, open loop from digital twin to hardware is a mid-horizon step toward shared experimental capacity — not a street-legal robotaxi timeline.
• Who should care: Academic AV labs, imitation-learning researchers, and groups that need reproducible sim-to-real driving without a fleet.

What the paper actually did

The paper presents a low-cost, open experimental platform for end-to-end autonomous driving on miniature Ackermann vehicles: a physical car, a printed urban track, data-collection tools, trajectory registration, and a Webots digital twin, so simulation methods can be run on hardware in a controlled way.

As a first baseline they implement command-conditioned behavior cloning: a neural policy takes an onboard camera image plus a high-level navigation command and outputs steering and speed. On the real vehicle in closed loop, the learned policy follows lanes and executes commanded turns with 6.1 cm mean cross-track error versus 4.7 cm for human demonstrations. In the twin, camera field of view is highly consequential: mean cross-track error falls from 35.6 cm to 3.3 cm when FOV widens from 58° to 120°. Using the twin for synthetic data plus a learned sim-to-real image translator, a higher-capacity policy trained on synthetic plus real demonstrations is the only setup that completes all four track routes; a compact baseline and the same larger net trained on real data alone complete fewer. They release the platform for reproducible research.

What makes this disruptive

The scarce resource is a closed-loop AV experiment you can afford to break. An open miniature stack that actually connects Webots to a real Ackermann car — and reports when sim-to-real data is load-bearing — pressures labs that only publish in simulation.

The FOV ablation (35.6 → 3.3 cm) is a teaching result as much as a methods result. Completing all four routes only with synthetic+real + a larger policy is a concrete sim-to-real lesson. This does not move public-road autonomy on a calendar.

Why it matters (outside the lab)

Abundance lens: reliable mobility research still needs scarce capital equipment. Lowering the cost of a real closed loop is a step toward shared experimental capacity rather than a few well-funded test tracks.

Horizon is mid-range for research defaults, long for public roads. Near-term: other groups can copy the platform. Medium-term: whether the baseline remains interesting matters less than whether the testbed stays maintained. No robotaxi year.

Limitations & open questions

The track is printed and miniature; 6.1 cm error is on that scale, not city streets. Only a behavior-cloning baseline is claimed as first. Route-completion results are for four track routes under the authors’ lighting and commands.

Preprint ≠ product. Abundance is not automatic: open hardware still has integration cost. Safety and legal autonomy are out of scope.