CARLA: An Open Urban Driving Simulator

Dec. 24, 2017, 5:03 a.m. By: Kirti Bakshi


CARLA(Car Learning To Act), a creation of Dosovitskiy and co. is an open-source simulator that is used for the purpose of autonomous driving research. CARLA has been developed from the ground up in order to support development, training, as well as validation of autonomous urban driving systems. In addition to open-source code and protocols that it presents itself to us with, the system also provides open digital assets like urban layouts, buildings, vehicles that were created for this purpose and can, therefore, be used freely.

CARLA is basically used to study the performance of three different approaches that are all related to autonomous driving:

  • a classic pipeline that is modular,

  • an end-to-end model that is trained through imitation learning,

  • an end-to-end model that is trained through reinforcement learning.

The simulation platform supports flexible specification of sensor suites as well as different environmental conditions. The evaluation of the approaches have been made in controlled scenarios of increasing difficulty, and their performance has been examined through metrics that have been provided by CARLA, thus illustrating the utility of the platform for autonomous driving research.

The system, called CARLA (Car Learning to Act), simulates a wide range of driving conditions and endlessly repeats dangerous situations that in return help learning. The team has already made use of it to evaluate the performance related to several different approaches to autonomous driving. The team created their own simulator CARLA that offers a library of assets that can further be arranged into towns under the various weather as well as lighting conditions. The library is comprised of 16 animated vehicle models, 40 different buildings, and 50 animated pedestrians to add to it.


The team made the use of these in order to create two towns with several kilometres of drivable roads and then further tested three different approaches to training self-driving systems. “The approaches are evaluated in controlled scenarios of increasing difficulty,” said the team.

The results then showed that the system can actually turn out to play a useful role. In the same context, the team has published a video of the resulting driving behaviour that not only clearly shows how well the systems can perform but also why the training of this kind cannot be done on real roads. Reason like: the cars sometimes drive on the sidewalk, on the opposite side of the road, hit other cars, and so on add to it.

Roadmap to Improvement:

There is a continuous work being done to improve CARLA, and the contributions from the community are also welcomed. The goals that are most immediate are as follows:

  • Releasing the methods that have been evaluated in the CARLA paper.

  • Addition of a LiDAR sensor.

  • Allowing flexible as well as user-friendly import and editing of maps along with it.

  • Providing the users with the benefit to control non-player characters and therefore set up user-specified scenarios.

It can be agreed that a system like CARLA can never replace the driving time on real roads but on the contemporary, it can also provide a useful as well as a safe testing ground for new ideas. And that is what makes it important.

CARLA is an open source system that is free to use for non-commercial purposes. So anyone who is interested can give it a go at its official site whose link is mentioned below. “We hope that CARLA will enable a broad community to actively engage in autonomous driving research,” says the team.

For More Information: GitHub

CARLA: An Open Urban Driving Simulator:

Video Source: Intel VCL