Racecar Gym

A gym environment for a miniature, F1Tenth-like racecar using the bullet physics engine with pybullet.

Introduction

This repository is for NTUST 2023 Reinforcement Learning in Human-Computer Interaction course - competition #1.\ This environment is originally from here.

Prerequisites

• Anaconda or other virtual env
• Python 3.10
• CPU or NVIDIA GPU
• Suitable PyTorch ## Installation

You can install racecar_gym with the following commands:

shell_script git clone https://github.com/MingCongSu/racecar_gym.git cd racecar_gym pip install -e .

Download Maps (Tracks)

Here is how you can do this from the command line: ```shell_script cd ./models/scenes

For Linux (including WSL)

wget https://github.com/MingCongSu/racecargym/releases/download/trainingtracks-v1/trainingtracks.zip unzip trainingtracks.zip

For Windows

wget -O trainingtracks.zip https://github.com/MingCongSu/racecargym/releases/download/trainingtracks-v1/trainingtracks.zip Expand-Archive -Path .\trainingtracks.zip -DestinationPath ./ `` After installation, go back toracecargymfolder and runtestenv.pyto test the environment: ``shellscript

go back to racecar_gym folder

python testenv.py `` There should beracecartest_env.mp4undervideos` folder

Environments

The observation and action space is a Dict holding the agents and their id's. The observation and action space for a single agent is also a Dict, which is described in more detail below. In general, observations are obtained through sensors and commands are executed by actuators. Vehicles can have multiple sensors and actuators. Those are described in the vehicle configuration (e.g. differential racecar). Agents, which consist of a vehicle and an assigned task, are specified in the scenario file (e.g. austria.yml). In this file, agents are described by the sensors to use (note that they must be available in the vehicle configuration) and the corresponding task. Have a look at tasks to see all available tasks.

Example: yaml world: name: austria agents: - id: A vehicle: name: racecar sensors: [lidar, pose, velocity, acceleration] actuators: [motor, steering] color: blue # default is blue, one of red, green, blue, yellow, magenta or random task: task_name: maximize_progress params: {laps: 1, time_limit: 120.0, terminate_on_collision: False}

This example specifies a scenario on the Austria track. One agent with id A is specified. The agent controls the differential drive racecar defined in differential racecar, identified by its name. The scenario tells the agent to use only the specified sensors (lidar, pose, velocity, acceleration). Optionally, one can also specify a color for the car. The default color is blue. Available colors are listed above.

The task which is assigned to this agent is also identified by name (implementations can be found in tasks). Task parameters are passed by the dict params.

Observations

Observations are obtained by (possibly noisy) sensors. Parameters for the sensors as well as the level of noise, can be configured in the corresponding vehicle configuration (e.g. differential racecar). In the scenario specification, one can specify which of the available sensors should be actually used. The observation space is a dictionary where the names of the sensors are the keys which map to the actual measurements. Currently, five sensors are implemented: pose, velocity, acceleration, LiDAR and RGB Camera. Further, the observation space also includes the current simulation time.

|Key|Space|Defaults|Description| |---|---|---|---| |pose|Box(6,)| |Holds the position (x, y, z) and the orientation (roll, pitch, yaw) in that order.| |velocity|Box(6,)| |Holds the x, y and z components of the translational and rotational velocity.| |acceleration|Box(6,)| |Holds the x, y and z components of the translational and rotational acceleration.| |lidar|Box(<scans>,)|scans: 1080|Lidar range scans.| |rgb_camera|Box(<height>, <width>, 3)|height: 240, width: 320|RGB image of the front camera.|

Actions

The action space for a single agent is a defined by the actuators of the vehicle. By default, differential racecar defines two actuators: motor and steering. The action space is therefore a dictionary with keys motor and steering.

Alternatevely, the agent can control the target speed and steering, but must be defined in the scenario specification. In this case, the action space is a dictionary with keys speed and steering.

Note, that the action space of the car is normalized between -1 and 1. The action space can include the following actuators:

| Key |Space| Description | |----------|---|-----------------------------------------------------------------------------| | motor |Box(low=-1, high=1, shape=(1,))| Throttle command. If negative, the car accelerates backwards. | | speed |Box(low=-1, high=1, shape=(1,))| Normalized target speed. | | steering |Box(low=-1, high=1, shape=(1,))| Normalized steering angle. |

State

In addition to observations obtained by sensors, the environment passes back the true state of each vehicle in each step (the state is returned as the info dictionary). The state is a dictionary, where the keys are the ids of all agents. Currently, the state looks like this:

|Key|Type|Description| |---|---|---| |wallcollision|bool|True if the vehicle collided with the wall.| |opponentcollisions|List[str]|List of opponent id's which are involved in a collision with the agent.| |pose|NDArray[6]|Ground truth pose of the vehicle (x, y, z, roll, pitch, yaw).| |acceleration|NDArray[6]|Ground truth acceleration of the vehicle (x, y, z, roll, pitch, yaw).| |velocity|NDArray[6]|Ground truth velocity of the vehicle (x, y, z, roll, pitch, yaw).| |progress|float|Current progress in this lap. Interval: [0, 1]| |time|float|Simulation time.| |checkpoint|int|Tracks are subdivided into checkpoints to make sure agents are racing in clockwise direction. Starts at 0.| |lap|int|Current lap.| |rank|int|Current rank of the agent, based on lap and progress.| |wrong_way|bool|Indicates wether the agent goes in the right or wrong direction.| |observations|Dict|The most recent observations of the agent.

Available API's

Gym API

To use the Gym API you can either instantiate environments with the standard keys or by loading custom scenarios. In either case, you have to load the gym_api module from this package: ```python import gymnasium import racecargym.envs.gymapi

For predefined environments:

env = gymnasium.make( id='SingleAgentAustria-v0', render_mode='human' )

For custom scenarios:

env = gymnasium.make( id='SingleAgentRaceEnv-v0', scenario='path/to/scenario', rendermode='rgbarrayfollow', # optional: 'rgbarraybirdseye' render_options=dict(width=320, height=240, agent='A') # optional )

done = False resetoptions = dict(mode='grid') obs, info = env.reset(options=resetoptions)

while not done: action = env.action_space.sample() obs, rewards, terminated, truncated, states = env.step(action) done = terminated or truncated

env.close() The predefined env-strings are of the form Agent-v0 e.g.: MultiAgentAustria-v0 ```

Maps

Currently available maps are listed below. The gridmaps are originally from the F1Tenth repositories.

| Image | Name | |---------------------------------------|----------| | | Austria | | | | Circle |

Notes

Please note that this is work in progress, and interfaces might change. Also more detailed documentation and additional scenarios will follow.