This repository contains the tensorflow implementation for training a reinforcement learning based map-less navigation model, as described in the paper:
Reinforced Imitation: Sample Efficient Deep Reinforcement Learning for Map-less Navigation by Leveraging Prior Demonstrations

1. Ubuntu
2. Python 2.7
3. ROS Indigo or ROS Kinetic
4. Stage-ros simulator, with add_pose_sub enabled. Can be found in this branch of the repository.

1. First run the stage simulator: roslaunch reinforcement_learning_navigation stage_sim.launch
2. In a separate terminal, run the training code: rosrun reinforcement_learning_navigation train_cpo.py --output_name $experiment_name$
   In order to use pre-trained weights from imitation learning, add the arguments --jump_start 1 --model_init $path_to_policy_weights$

If you use our code in your research, please cite our paper.

@ARTICLE{pfeiffer2018ral,
author={M. Pfeiffer and S. Shukla and M. Turchetta and C. Cadena Lerma and A. Krause and R. Siegwart and J. Nieto},
journal={IEEE Robotics and Automation Letters},
title={{Reinforced Imitation: Sample Efficient Deep Reinforcement Learning for Map-less Navigation by Leveraging Prior Demonstrations}},
year={2018},
volume={3},
number={4},
pages={4423-4430}
}

Our training model uses Constrained Policy Optimization : [Paper] [Code]