Learning to navigate in five minutes
A brief implementation of DAgger in multi-agent environment. It Only takes five minutes to learn a basic policy from human-beings.
- Forward looking single line lidar
- Relative target pose
- [Acc, Dec, Left_turn_slightly, Right_turn_slightly]
- rospy
- gazebo
- Pytorch
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
The green robot is controlled by DAgger framework(human and network), the red robot is controlled only by network.