Airsim as an Openai Gym env

I have been successful in establishing a connection with aimsun by running the baseline_racer after running airsim. After this, the drone flies in some trajectory and then crashes or possibly wanders. I wanted to sample the action space as in any gym environment or manually supply actions for the drone for testing purposes.

How can I integrate the following functionality in the code-

import gym
env = gym.make('CartPole-v0') # possibly start the communication via the client here for Airsim
env.reset()
for _ in range(1000):
    env.step(env.action_space.sample()) # take a random action
env.close() 

I want to manually supply action so that once I know how the actions are being, the type and properties of the action can I replace the random sampling of action space with a Refined piece of Reinforcement learning snippet or model. Help would be appreciated.

[Deepakgthomas]

Sounds good. Thank you.

On Mon, Feb 1, 2021 at 4:08 PM Yann Bouteiller [e-mail] wrote:

@yannbouteiller https://github.com/yannbouteiller Why do you say that the integration of Airsim and OpenAI gym will work with tier 2. Could you please elaborate?

Thanks!

Well it actually works on all tiers, but tiers 1 and 3 are multi-agent so it comes with additional difficulties.

I have not announced it because I work on other things these days, the project in on hold and not in a presentable state, but I have silently open sourced this : https://github.com/yannbouteiller/gym-airsimdroneracinglab, where basically I have implemented a OpenAI Gym environment for multi-agent tiers as well. You may find it interesting.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/microsoft/AirSim-NeurIPS2019-Drone-Racing/issues/76#issuecomment-771192651, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFHXYWW7UAATFOVS6XYW2LTS44Q5DANCNFSM4I633MGQ .

[yannbouteiller]

@yannbouteiller Why do you say that the integration of Airsim and OpenAI gym will work with tier 2. Could you please elaborate?

Thanks!

Well it actually works on all tiers, but tiers 1 and 3 are multi-agent so it comes with additional difficulties.

I have not announced it because I work on other things these days, the project in on hold and not in a presentable state, but I have silently open sourced this : https://github.com/yannbouteiller/gym-airsimdroneracinglab, where basically I have implemented a OpenAI Gym environment for multi-agent tiers as well. You may find it interesting.