carla/Docs/benchmark_structure.md

67 lines
2.3 KiB
Markdown

Driving Benchmark Structure
-------------------
The figure below shows the general structure of the driving
benchmark module.
![Benchmark_structure](img/benchmark_diagram.png)
>Figure: The general structure of the agent benchmark module.
The *driving benchmark* is the module responsible for evaluating a certain
*agent* in an *experiment suite*.
The *experiment suite* is an abstract module.
Thus, the user must define its own derivation
of *experiment suite*. We already provide the CoRL2017 suite and a simple
*experiment suite* for testing.
The *experiment suite* is composed by set of *experiments*.
Each *experiment* contains a *task* that consists of a set of navigation
episodes, represented by a set of *poses*.
These *poses* are tuples containing the start and end points of an
episode.
The *experiments* are also associated with a *condition*. A
condition is represented by a [carla settings](carla_settings.md) object.
The conditions specify simulation parameters such as: weather, sensor suite, number of
vehicles and pedestrians, etc.
The user also should derivate an *agent* class. The *agent* is the active
part which will be evaluated on the driving benchmark.
The driving benchmark also contains two auxiliary modules.
The *recording module* is used to keep track of all measurements and
can be used to pause and continue a driving benchmark.
The [*metrics module*](benchmark_metrics.md) is used to compute the performance metrics
by using the recorded measurements.
Example: CORL 2017
----------------------
We already provide the CoRL 2017 experiment suite used to benchmark the
agents for the [CoRL 2017 paper](http://proceedings.mlr.press/v78/dosovitskiy17a/dosovitskiy17a.pdf).
The CoRL 2017 experiment suite has the following composition:
* A total of 24 experiments for each CARLA town containing:
* A task for going straight.
* A task for making a single turn.
* A task for going to an arbitrary position.
* A task for going to an arbitrary position with dynamic objects.
* Each task is composed of 25 poses that are repeated in 6 different weathers (Clear Noon, Heavy Rain Noon, Clear Sunset, After Rain Noon, Cloudy After Rain and Soft Rain Sunset).
* The entire experiment set has 600 episodes.
* The CoRL 2017 can take up to 24 hours to execute for Town01 and up to 15
hours for Town02 depending on the agent performance.