Multi-robot Reinforcement Learning Scalable Training School (MRST) is a training and evaluation platform for reinforcement learning reasearch.

Check out the background and project goals in our paper "From Simulation to Reality: A Scalable Emulation Platform for Multi-robot Reinforcement Learning".

Our project require the installation of the following third-parties before getting started:

• Ubuntu 18
• python 3.7+
• torch 1.8+
• Ros Knetic or Melodic
• Turtlebot3

Create a workspace, clone the repo using git command:

git clone https://github.com/Blackmamba-xuan/MRST.git

Move the folder mrst_simulation to your catkin workspace and build for it

cd ~/catkin_ws/src/
catkin_make

We implement several high impact reinforcement learning algorithms and extend them to multi-robot areas.

• Indepent Q-learning
• MADDPG
• COMA
• MAAC
• MASAC
• IPPO
• MAPPO
• VDN
• QMIX
• Hierarchical RL
• Others

Use the roslaunch command to run the simulator:

# launch a crosslane environment
roslaunch mrst_simulation turtlebot3_autorace_crossLane.launch

# launch a roundabout environment
roslaunch mrst_simulation turtlebot3_autorace_roundabout.launch

from Env import Env
   def main():
      env=Env(scenario="roundabout")
      n_episodes = 100
      n_agents=12
      episode_length=15
      for e in range(n_episodes):
         env.reset()
         for et_i in range(episode_length):
               print(et_i)
               actions=[[1] for i in range(n_agents)]
               next_obs, rewards, dones, speeds = env.step(actions, isTeamReward=True)
   if __name__ == "__main__":
      main()

For more detalis, please check our documentaion.