This repository uses Transfer Learning (TL) based approach to reduce on-board computation required to train a deep neural network for autonomous navigation via Deep Reinforcement Learning for a target algorithmic performance. A library of 3D realistic meta-environments is manually designed using Unreal Gaming Engine and the network is trained end-to-end. These trained meta-weights are then used as initializers to the network in a real test environment and fine-tuned for the last few fully connected layers. Variation in drone dynamics and environmental characteristics is carried out to show robustness of the approach. The repository containing the code for simulated environment on a simulated drone can be found @ DRLwithTL-Sim

The current version of DRLwithTL-Sim supports Windows and requires python3. It’s advisable to make a new python virtual environment for this project and install the dependencies. Following steps can be taken to download get started with DRLwithTL-Real

git clone https://github.com/aqeelanwar/DRLwithTL_real.git

The provided requirements.txt file can be used to install all the required packages. Use the following command

cd DRLwithTL_real
pip install –r requirements.txt

This will install the required packages in the activated python environment.

Once you have the required packages, you can take the following steps to run the code

Download the imagenet weights for the network from the here and place it in

DeepNet/models/imagenet.npy

1. Turn DJI Tello On
2. Connect your workstation with DJI Tello over Wi-Fi

The RL parameters for the DRL algorithm can be set using the provided config file and are self-explanatory.

cd DRLwithTL_real\configs
notepad config.cfg (#for windows)

The DRL code can be started using the following command

cd DRLwithTL
python main_code.py

While the simulation is running, RL parameters such as epsilon, learning rate, average Q values and loss can be viewed on the tensorboard. The path depends on the env_type, env_name and train_type set in the config file and is given by 'DeepNet/models/<run_name>/<env_type>/'. An example can be seen below

cd DeepNet/models/Tello_indoor/VanLeer/
tensorboard --logdir agent

Pygame screen can be used to control the DJI Tello drone on the fly. Following control keys are supported

1. 'M' : Top toggle between automatic and manual mode. In automatic phase, the drone uses RL training to navigate around the environment using epsilon greedy algorithm. In manual phase (default starting phase) the user has control over the drone and can use left, right, up, down, w, a, s, d keys to navigate. All the other control keys mentioned below only works when the drone is in the manual mode
2. 'Escape': Quits the code
3. 'L' : Save the DNN weights and Replay memory to the path specified in the config file.

If you find this repository useful for your research please use the following bibtex citations

@ARTICLE{2019arXiv191005547A,
       author = {{Anwar}, Aqeel and {Raychowdhury}, Arijit},
        title = "{Autonomous Navigation via Deep Reinforcement Learning for Resource Constraint Edge Nodes using Transfer Learning}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Machine Learning, Statistics - Machine Learning},
         year = "2019",
        month = "Oct",
          eid = {arXiv:1910.05547},
        pages = {arXiv:1910.05547},
archivePrefix = {arXiv},
       eprint = {1910.05547},
 primaryClass = {cs.LG},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2019arXiv191005547A},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

@article{yoon2019hierarchical,
  title={Hierarchical Memory System With STT-MRAM and SRAM to Support Transfer and Real-Time Reinforcement Learning in Autonomous Drones},
  author={Yoon, Insik and Anwar, Malik Aqeel and Joshi, Rajiv V and Rakshit, Titash and Raychowdhury, Arijit},
  journal={IEEE Journal on Emerging and Selected Topics in Circuits and Systems},
  volume={9},
  number={3},
  pages={485--497},
  year={2019},
  publisher={IEEE}
}

• Aqeel Anwar - Georgia Institute of Technology

This project is licensed under the MIT License - see the LICENSE.md file for details