CARLA Benchmark

Running the Benchmark

The "carla" api provides a basic benchmarking system, that allows making several tests on a certain agent. We already provide the same benchmark used in the CoRL 2017 paper. By running this benchmark you can compare the results of your agent to the results obtained by the agents show in the paper.

Besides the requirements of the CARLA client, the benchmark package also needs the future package

$ sudo pip install future

By running the benchmark a default agent that just go straight will be tested. To run the benchmark you need a server running. For a default localhost server on port 2000, to run the benchmark you just need to run

$ ./


$ python

Run the help command to see options available

$ ./ --help

Benchmarking your Agent

The benchmark works by calling three lines of code

corl = CoRL2017(city_name=args.city_name, name_to_save=args.log_name)
agent = Manual(args.city_name)
results = corl.benchmark_agent(agent, client)

This is excerpt is executed in the example.

First a benchmark object is defined, for this case, a CoRL2017 benchmark. This is object is used to benchmark a certain Agent.
On the second line of our sample code, there is an object of a Manual class instanced. This class inherited an Agent base class that is used by the benchmark object. To be benchmarked, an Agent subclass must redefine the run_step function as it is done in the following excerpt:

def run_step(self, measurements, sensor_data, target):
    Function to run a control step in the CARLA vehicle.
    :param measurements: object of the Measurements type
    :param sensor_data: images list object
    :param target: target position of Transform type
    :return: an object of the control type.
    control = VehicleControl()
    control.throttle = 0.9
    return control

The function receives measurements from the world, sensor data and a target position. With this, the function must return a control to the car, i.e. steering value, throttle value, brake value, etc.

The measurements, target, sensor_data and control types are described on the documentation.

Creating your Benchmark

Tutorial to be added