CARLA Gym Wrapper

A simple gym environment wrapping Carla, a simulator for autonomous driving research. The environment is designed for developing and comparing reinforcement learning algorithms. Trackable costs also enable the application of safe reinforcement learning algorithms.

carla

Features

rendering
weather
different observation types (state, pixel)
traffic
autopilot
vehicle and map selection
configurable environments
collect, save and load data
costs (for safe RL)

Observation Space

Pixel: (3, 84, 84)

Index	Value
0	r channel 84 x 84
1	g channel 84 x 84
2	b channel 84 x 84

State: (9, )

Index	Value
0	x_pos
1	y_pos
2	z_pos
3	pitch
4	yaw
5	roll
6	acceleration
7	angular_velocity
8	velocity

Action Space

Action: (2, )

Index	Value	Min	Max
0	throttle_brake	-1	1
1	steer	-1	1

CARLA Setup

Add the following environment variables:

export PYTHONPATH=$PYTHONPATH:/opt/carla-simulator/PythonAPI
export PYTHONPATH=$PYTHONPATH:/opt/carla-simulator/PythonAPI/carla/
export PYTHONPATH=$PYTHONPATH:/opt/carla-simulator/PythonAPI/carla/dist/carla-0.9.9-py3.7-linux-x86_64.egg

Install the following extra libraries

pip install pygame
pip install networkx
pip install dotmap
pip install gym

Open a new terminal session, and run the CARLA simulator:
```
bash CarlaUE4.sh -fps 20
```

Configure environments

Open: carla_env/_init_.py

Insert a new environment configuration

register(
id='CarlaEnv-state-town01-v1',
entry_point='carla_env.carla_env:CarlaEnv',
max_episode_steps=500,
kwargs={
    'render': True,
    'carla_port': 2000,
    'changing_weather_speed': 0.1,
    'frame_skip': 1,
    'observations_type': 'state',
    'traffic': True,
    'vehicle_name': 'tesla.cybertruck',
    'map_name': 'Town01',
    'autopilot': True
}
)

Environment usage

import gym
import carla_env

env = gym.make('CarlaEnv-pixel-v1')
env.reset()
done = False
while not done:
    action = [1, 0]
    next_obs, reward, done, info = env.step(action)
env.close()

Example script:

python carla_env_test.py

Data collection

How to wrap the environment with the data collection wrapper:

env = gym.make('CarlaEnv-pixel-v1')
env = DataCollector(env, steps=200, save_dir='./output')

Load existing dataset:

env = DataCollector(env, steps=200, save_dir='./output', load_dir='./output/dataset_200.pkl')