BeamNG.drive/tech based simulation environments.

1. Obtain and install BeamNG.tech or BeamNG.drive (e.g. from Steam) - https://www.beamng.com/game/
2. Create the intended user path, e.g. /beamng_workspace/
3. Inside that directory create ResearchHelper.txt
4. Create a mods directory with the version of BeamNG, eg. /beamng_workspace/0.25/mods/
5. Put the BeamNGpy.zip file included in this repo into the directory created in step 4.
6. Setup a python environment, e.g using Conda:

   conda create -y --name beamng-envs python=3.8
   conda activate beamng-envs

7. Install requirements:

   pip install -r requirements.txt

Future details on setting up and using BeamNG via its Python API can be found here https://github.com/BeamNG/BeamNGpy

from beamng_envs.bng_sim.beamngpy_config import BeamNGPyConfig
from beamng_envs.envs.track_test.track_test_param_space import TRACK_TEST_PARAM_SPACE_GYM
from beamng_envs.envs.track_test.track_test_env import TrackTestEnv
from beamng_envs.envs.track_test.track_test_config import TrackTestConfig

# Set paths are required
track_test_config = TrackTestConfig(
    output_path='track_test_results',
    bng_config=BeamNGPyConfig(
        home='T:/Program Files (x86)/Steam/steamapps/common/BeamNG.drive',
        user="C:/beamng_workspace/"
    )
)
param_set = TRACK_TEST_PARAM_SPACE_GYM.sample()
env = TrackTestEnv(params=param_set, config=track_test_config)
results, history = env.run()

There are a number of scripts that demo running an environment and collecting the result to MLflow. The input args are the same for each, and require the BeamNG paths to be set.

The MLflow logging shown in the examples is optional; full results are saved to disk in the output directory in json format.

Sets a car config for the Scintila Rally car, use an AI to drive a lap of the track, and record lap time and sensor data.

python -m scripts.run_batch_track_tests -N 5 --beamng_path /SteamLibrary/steamapps/common/BeamNG.drive --beamng_user_path /beamng_workspace/

Searches for all available car templates, selects one, then drive it into a barrier in the destruction section of grid map v2 at a set speed. Records sensor data.

python -m scripts.run_batch_crash_tests -N 5 --beamng_path /SteamLibrary/steamapps/common/BeamNG.drive --beamng_user_path /beamng_workspace/

Selects a totally random part config for the Hirochi Sunburst car, then races it down the drag strip in grid map v2. Records time, speed, etc. along with sensor data.

python -m scripts.run_batch_drag_strip -N 5 --beamng_path /SteamLibrary/steamapps/common/BeamNG.drive --beamng_user_path /beamng_workspace/

It's possible to run environments in multiple BeamNG instances simultaneously. See run_batch_parallel_crash_tests.py for an example.

The DiskResults class can be used to load saved results from an environment output folder or from an on disk MLflow experiment. It automates loading the saved json data into Pandas formats.

Foe example to load the results from an experiment and to access the timeseries sensor data and other scalar params/results:

from beamng_envs.data.disk_results import DiskResults

# Load from the environment output path = this can be passed from an existing env directly, using
# env.disk_results.output_path, or as disk path if the evn is no longer in memory.
full_results = DiskResults.load(path='track_test_results/[uuid for result]')

# Access the timeseries sensor data (pd.DataFrame)
timeseries_df = full_results.ts_df

# Access the scalar params/results (pd.Series)
scalars = full_results.scalars_series

See run_single_track_test.py for a more complete example.

mlflow ui -h 0.0.0.0 -p 5555

http://0.0.0.0:5555

Each version of the Beamng python api supports specific versions of Beamng - see https://github.com/BeamNG/BeamNGpy#compatibility

The environments here are compatible with the following versions: