Skip to content

Latest commit

 

History

History
244 lines (197 loc) · 7.75 KB

README.md

File metadata and controls

244 lines (197 loc) · 7.75 KB
Raw

gym-airsim

OpenAI Gym-compatible environments of AirSim for multirotor control in RL problems. All gym environments have corresponding Unreal Engine environments that are provided in the release section ready for use (Linux only).

Paper

The work in this repository is part of a publication made in the IEEE ICME 2020.
More technical details are available in Leveraging Deep Reinforcement Learning For Active Shooting Under Open-World Setting (IEEE ICME 2020). Please cite this as:

@INPROCEEDINGS{9102966,  
  author={A. {Tzimas} and N. {Passalis} and A. {Tefas}},  
  booktitle={2020 IEEE International Conference on Multimedia and Expo (ICME)},   
  title={Leveraging Deep Reinforcement Learning For Active Shooting Under Open-World Setting},   
  year={2020},  
  volume={},  
  number={},  
  pages={1-6},
}

Unreal Engine Environments

The Unreal Engine project provided in the releases section constists of 3 environments. Some simple commands for the UE console are provided. Launch the project and press the back quote/tilde button(~) on your keyboard to open the console and type one of the following (without quotes):

  • "open MultiHumanCity" : Opens the MultiHumanCity environment map.
  • "open StaticHuman" : Opens the StaticHuman environment map.
  • "open StaticCube" : Opens the StaticCube environment map.
  • "fullscreen" : Toggles between fullscreen and windowed mode.
  • "exit" : Exits the simulator.

MultiHumanCity

This environment represents a small part of a city. It consists of four buildings, bus stops, trees and other small items. This city includes 9 different 3D models of humans with a variety of clothing and facial features. The humans have a small idling animation.

city screenshot


StaticHuman

This environment consists of a square platform and the 3D model of a human in its center. The background is a simple skybox.

human screenshot


StaticCube

This environment consists of a square platform with a red cube in its center. The background is a simple skybox. Really simple environment for starting off with tracking tasks and AirSim.

cube screenshot


Gym Environments

FrontalShotDroneOnlyMulti

Corresponding Unreal Engine Environmet: MultiHumanCity

Description: The agent has to move the drone in front of the person's face in order to get a frontal close-up shot. In this environment there are 9 different people.

Observation: A (200)x(200) color image taken from the bottom_center camera of the AirSim drone. The resolution of the observation image is determined in the AirSim settings.

Actions: Type: Discrete(9)

  • 0 Drone forwards
  • 1 Drone backwards
  • 2 Drone left
  • 3 Drone right
  • 4 Drone up
  • 5 Drone down
  • 6 Drone yaw left
  • 7 Drone yaw right
  • 8 Do nothing

Reward: The reward depends on the drone's distance from the frontal position and the angle between the target's direction and the drone's look direction.

Starting State: The drone is placed in a random position in front of a random target human facing towards him.

Episode Termination:

  • The person's face moves out of the camera frame.
  • The drone moves far away from the person's face.
  • The drone collides with another object.
  • 20 seconds pass from the beggining of the episode.


FrontalShotContMulti

Corresponding Unreal Engine Environmet: MultiHumanCity

Description: The agent has to move the drone in front of the person's face in order to get a frontal close-up shot. In this environment there are 9 different people.

Observation: A (200)x(200) color image taken from the front_center camera of the AirSim drone. The resolution of the observation image is determined in the AirSim settings.

Actions: Type: Box(5)

  • 0 Drone forwards/backwards
  • 1 Drone left/right
  • 2 Drone up/down
  • 3 Camera tilt up/down
  • 4 Camera pan left/right

Reward: The reward depends on the drone's distance from the frontal position and the angle between the target's direction and the drone's look direction.

Starting State: The drone is placed in a random position in front of a random target human facing towards him.

Episode Termination:

  • The person's face moves out of the camera frame.
  • The drone moves far away from the person's face.
  • The drone collides with another object.
  • 20 seconds pass from the beggining of the episode.


FrontalShotDroneOnly

Corresponding Unreal Engine Environmet: StaticHuman

Description: The agent has to move the drone in front of the person's face in order to get a frontal close-up shot.

Observation: A (200)x(200) color image taken from the bottom_center camera of the AirSim drone. The resolution of the observation image is determined in the AirSim settings.

Actions: Type: Discrete(9)

  • 0 Drone forwards
  • 1 Drone backwards
  • 2 Drone left
  • 3 Drone right
  • 4 Drone up
  • 5 Drone down
  • 6 Drone yaw left
  • 7 Drone yaw right
  • 8 Do nothing

Reward: The reward depends on the drone's distance from the frontal position and the angle between the target's direction and the drone's look direction.

Starting State: The drone is placed in a random position in front of the target human facing towards him.

Episode Termination:

  • The person's face moves out of the camera frame.
  • The drone moves far away from the person's face.
  • The drone collides with another object.
  • 10 seconds pass from the beggining of the episode.


FrontalShotCont

Corresponding Unreal Engine Environmet: StaticHuman

Description: The agent has to move the drone in front of the person's face in order to get a frontal close-up shot.

Observation: A (200)x(200) color image taken from the bottom_center camera of the AirSim drone. The resolution of the observation image is determined in the AirSim settings.

Actions: Type: Box(5)

  • 0 Drone forwards/backwards
  • 1 Drone left/right
  • 2 Drone up/down
  • 3 Camera tilt up/down
  • 4 Camera pan left/right

Reward: The reward depends on the drone's distance from the frontal position and the angle between the target's direction and the drone's look direction.

Starting State: The drone is placed in a random position in front of the target human facing towards him.

Episode Termination:

  • The person's face moves out of the camera frame.
  • The drone moves far away from the person's face.
  • The drone collides with another object.
  • 10 seconds pass from the beggining of the episode.


FollowTargetCamOnly

Corresponding Unreal Engine Environmet: StaticCube

Description: The drone performs a random trajectory and the agent has to keep track of the red cube using only the drone's camera.

Observation: A (256)x(144) color image taken from the bottom_center camera of the AirSim drone. The resolution of the observation image is determined in the AirSim settings.

Actions: Type: Discrete(5)

  • 0 Do nothing
  • 1 Tilt drone camera up
  • 2 Tilt drone camera down
  • 3 Pan drone camera left
  • 4 Pan drone camera right

Reward: Reward depends on the distance of the red cube from the center of the camera's frame.

Starting State: Drone starts in a random position in the air with the camera pointed at the direction of the red cube.

Episode Termination:

  • Cube moves out of the camera frame.
  • Drone completes the trajectory and stops moving.


Acknowledgement

This work was supported by the European Union's Horizon 2020 Research and Innovation Program (OpenDR Project ) under Grant 871449. This publication reflects the authors' views only. The European Commission is not responsible for any use that may be made of the information it contains.