A virtual autonomous car has learned to consider several possible reactions from nearby human drivers before deciding how to accelerate or change lanes.
On a real road, two motorists placed in a similar situation do not necessarily react in the same way. One may let a car pass, while another may quickly merge in front of it. For an automated system, relying solely on current positions and speeds therefore limits its ability to prepare its next action.
Pixabay illustrative image
The researchers designed a method that does not attempt to predict a single future. Instead, it produces several plausible trajectories for nearby vehicles based on their movements and interactions. These possibilities then become information that can be used by the program responsible for driving. In other words, it can prepare its decision by taking several possible human behaviors into account.
This anticipation also helps prevent the AI from too often exploring unrealistic actions during training. The system learns through successive trials in a simulation, receiving penalties in the event of a collision or unnecessary maneuver. It also receives a reward when it makes efficient progress without sacrificing safety.
To make the trials more realistic, the team used trajectories recorded on German highways from the HighD dataset. Drones had tracked the movements of vehicles from above. In this study, 2,325 real trajectories were used in particular to train the module responsible for anticipating movements.
The results are highly encouraging in this virtual environment. In the reference scenario, the collision rate reaches 0.1%, compared with 0.7% to 2.7% for four machine-learning methods used for comparison. During emergency braking, it reaches 1.0%, compared with about 24% for two competing methods tested under the same conditions.
The team also tested traffic made 50% denser. The collision rate reached 0.3%, compared with 1.8% for the two selected comparison methods. The experiment also showed a higher average speed. The authors attribute this result in particular to the ability to anticipate several possible developments in traffic before taking action.
One step remains necessary before considering real-world use: the system will need to learn to adapt to situations absent from its initial data. The researchers also plan to incorporate energy consumption and extend their method to multiple autonomous vehicles cooperating through direct communication.