README.md

Coupled-Imitation-and-Reinforcement-Learning

Code implementation for tightly and effiiciently coupling imitation learning and reinforcement learning to derive mobile robot goal-driven navigation under minimum demonstrated data, meanwhile achieving better sample efficiency and training safety.

Required software ubuntu 14.04 ros-indigo gazebo 7.0

Instruction

1)To spawn the simulation environment in customized gazebo world, run the following command.

sudo cp -r ~/catkin_ws ./ cd catkin_ws && catkin_make

There may require to install some packages specified in the warning info with the following command

sudo apt-get install ros-kinetic-*

2)To run the Reinforcement learning model training

Record the human demonstrated navigation trajectories

cd ~/UGV_navigation_RL/record_action && python record_action.py

3)Pretrain the model with Bootstrapping Imitaion Learning (BIM)

cd ~/UGV_navigation_RL/offline_training && python offline_training.py

4)Training the model with RL combining Near-Optimal-Policy(NOP) strategy

cd ~/UGV_navigation_RL/train_human_exp && python train_human_exp1.py

5)Transfer the model to the husky-A200 (differential drive mobile robot)

cd ~/UGV_navigation_RL/real_test && python train_human_exp.py

Main results

Training Curve

Performance Video (click the diagram below for playing, the diagram shows us the framework of CIRL)

![CIRL_VIDEO]