Research paper

Can a basic sensor suite drive?

Testing camera-plus-radar autonomous driving in a realistic simulator of a production car

Owen Dibble · November 2026

I modeled a 2025 Kia K4 GT-Line's driver-assist sensors in CARLA and asked whether a model trained on only those sensors can drive.

The CARLA simulator running the K4: radar top-down view, main view, rear camera, and the forward camera with detections.
Figure 1. The CARLA simulator running the K4: radar top-down view, main view, rear camera, and the forward camera with detections.

01

Abstract

Most new cars ship with a basic driver-assist sensor suite: one forward camera and a few radars. No lidar. This project asks whether that hardware alone could support autonomous driving.

I tested the question in CARLA 1.0 using a simulated 2025 Kia K4 GT-Line. Sensor realism came first. If the simulated sensors are cleaner than the real ones, the answer means nothing.

02

The simulated car and sensors

The car carries the same kind of sensors a production K4 has, and nothing more.

  • Forward camera at 1600×900 with a 100° field of view.
  • Front radar with a long-range scan and a wide short-range scan.
  • Two rear-corner blind-spot radars.
  • No lidar and no ultrasonics.

Real radar is messy, so the simulated radar is too. Each radar adds noise, drops detections, caps how many it reports, and produces ghost detections that aren't really there.

Figure · Sensor layout diagram

Figure 2. Where each sensor sits on the car and what it covers.

Figure · Camera view with detections

Figure 3. The forward camera view with detected vehicles and lane lines drawn on top.

Figure · Radar top-down view

Figure 4. Radar detections from above, colored by distance. Rings are every 10 meters.

03

The expert driver (teacher)

The expert is a rule-based driver that can see the simulator's ground truth. It knows exactly where the lanes, cars, and signs are. It does two jobs: it drives, and it labels the data the student learns from.

  • Lane following.
  • Smooth braking.
  • Stopping at stop lines and stop signs.
  • Routing through junctions.
  • Safe lane changes.

04

The student model

The student is a behavior-cloning network. It only gets what the car's real sensors would give it: the simulated camera, the radar returns, and its own speed.

It outputs steering, throttle, and brake. It also learns to predict lane error and heading error as extra training targets, which pushes it to understand where it sits in the lane instead of just copying controls.

Figure · Expert vs student driving

Figure 5. The same route driven by the expert and by the student model.

05

Experiments and results

Testing is underway. I'll post the numbers once they're measured.

Route completion

Coming soon

Share of each test route finished.

Collisions

Coming soon

Collisions per kilometer driven.

Lane keeping

Coming soon

Average distance from lane center.

Stop compliance

Coming soon

Stop lines and signs handled correctly.

Figure · Results charts

Figure 6. Charts comparing the student model against the expert across test routes.

06

Limitations

  • Testing started on a single map.
  • The expert uses privileged information the real car would never have.
  • Some map glitches needed manual fixes.
  • Transfer from the simulator to a real car is untested.

07

Next steps

  • Scored test routes to measure driving quality.
  • A correction loop where the expert fixes the model's mistakes and the model retrains on them.
  • More maps.
  • Sensor-failure tests.

Send me a message

Have a question or an opportunity? I'd love to hear from you.

Start a conversation

© 2026 Owen Dibble

St. Louis, Missouri