multiDAgger

Learning to navigate in five minutes

Overview

A brief implementation of DAgger in multi-agent environment. It Only takes five minutes to learn a basic policy from human-beings.

Inputs:

• Forward looking single line lidar
• Relative target pose

Output:

• [Acc, Dec, Left_turn_slightly, Right_turn_slightly]

Setup

Dependencies

• rospy
• gazebo
• Pytorch

Commands

Start the simulation environment:

roscore
rosrun gazebo_ros gazebo multiDAgger.world

Use the turtlesim teleop_key as remote controller:

rosrun turtlesim turtle_teleop_key

Start main loop (for training):

python referee.py

Start main loop (just for policy test):

python test_referee.py

Demo

The green robot is controlled by DAgger framework(human and network), the red robot is controlled only by network.