As a virtual lorry approaches from behind, the left grip of a stationary bicycle starts to vibrate. Researchers at Osnabrück University of Applied Sciences in Germany used Grand Theft Auto V (GTA V) to see whether cyclists could pick up that warning without looking down.
In a separate 2017 study, researchers in South Korea used images from GTA V to train artificial intelligence to predict when a car was about to collide.
The two studies tested ways to detect or communicate danger in a simulation, leaving open whether either approach could prevent crashes on public roads.
Testing Cyclist Alerts
The German team published its findings in March 2023 after fitting an ordinary bicycle with sensors that let volunteers steer, pedal and brake in the game.
The researchers modified GTA V to control the test conditions and chose a straight stretch of road where lorries regularly overtook riders on the left.
The grip began to vibrate when a lorry was 80 metres away, vibrated more strongly at 40 metres and was followed by a beep at 20 metres. A second prototype showed the approaching vehicle on a smartphone mounted on the bicycle, using coloured bars to indicate its distance and which side it was on.
The four volunteers preferred the vibrating grip because they could keep looking ahead. One rider wondered whether the signal would still be noticeable over cobblestones, where the surface itself would send vibrations through the bike.
Riders could turn the handlebars but could not lean into bends, and the researchers noticed a delay between steering and seeing the game respond.
The pilot established which alert the four volunteers preferred under those conditions, while detection on an actual road surface remained untested in the study.
Predicting Crashes
At the Korea Advanced Institute of Science and Technology, researchers generated traffic scenes by making nearby vehicles drive erratically while the player's car kept moving at a steady speed.
Their dataset contained 26,705 images taken just before collisions and 27,325 from scenes without a crash.
The researchers trained the software to judge from a single image whether a collision was imminent. Its best model classified images in the game-based test with 96.2% accuracy, compared with 89.6% for the strongest rules-based method the team evaluated.
The researchers found that the model picked up on details such as which way a nearby vehicle was pointing and how its wheels were angled.
The crash images used in the test came from the final half-second before impact, and the 2017 experiment left open how well the model would work with real traffic footage.
Limits of the Simulation
A separate 2018 experiment at Delft University of Technology used GTA V and eye-tracking equipment to see where 27 riders looked as they encountered simulated hazards.
Those situations included a car pulling out of a parking space and a car door opening in their path.
For the handlebar warning, the practical question is whether riders can feel it through the vibration of a real road.
For the AI, it is whether a warning would come early enough for a driver or vehicle to respond safely.
Frequently Asked Questions
- What was the purpose of the German study using GTA V?The study aimed to test cyclist alert systems using a vibrating grip and smartphone display to warn of approaching vehicles.
- What did the South Korean study focus on?The study focused on training AI to predict car collisions using images from GTA V.
- What were the limitations of these studies?The studies were conducted in simulations, leaving open questions about their effectiveness in real-world scenarios.
