racing

Racing

This example illustrates the training of an ego agent to drive, as fast and as far as possible, in traffic using the DreamerV2 reinforcement-learning algorithm.

The ego agent earns reward for the distance travelled per-step and is penalised for colliding with other vehicles and for going off-road.

Trained agent racing in traffic

Observation space

• Topdown RGB image
  • size (width x height): 64 pixels x 64 pixels
  • resolution: 1 meter/pixel

observation_space = gym.spaces.Dict({
    gym.spaces.Box(low=0, high=255, shape=(64,64,3), dtype=np.uint8),
})

Action space

• Throttle: [0,1]
• Brake: [0,1]
• Steering: [-1,1]

action_space = gym.spaces.Box(low=-1.0, high=1.0, shape=(3,), dtype=np.float32)

Setup

$ git clone https://github.com/huawei-noah/SMARTS.git
$ cd <path>/SMARTS/examples/rl/racing
$ python3.7 -m venv ./.venv
$ source ./.venv/bin/activate
$ pip install --upgrade pip
$ pip install -e .

Train

1. Train

   $ cd <path>/SMARTS/examples/rl/racing
   $ python3.7 run.py 

2. The trained model is saved into <path>/SMARTS/examples/rl/racing/logs/<folder_name> folder.

Evaluate

1. Evaluate

   $ cd <path>/SMARTS/examples/rl/racing
   $ scl envision start -s ./scenarios &
   $ python3.7 run.py --mode=evaluate --logdir="<path>/SMARTS/examples/rl/racing/logs/<folder_name>" --head

2. Go to localhost:8081 to view the simulation in Envision.

Docker

1. Train a model inside docker

   $ cd <path>/SMARTS
   $ docker build --file=<path>/SMARTS/examples/rl/racing/Dockerfile --network=host --tag=racing <path>/SMARTS
   $ docker run --rm -it --network=host --gpus=all racing
   (container) $ cd /src/examples/rl/racing
   (container) $ python3.7 run.py