Crossgap_IL_RL is the open-soured project of our IROS_2019 paper "Flying through a narrow gap using neural network: an end-to-end planning and control approach" (our preprint version on arxiv, our video on Youtube) , including some of the training codes, pretrain networks, and simulator (based on Airsim).
Introduction: Our project can be divided into two phases,the imitation learning and reinforcement learning. In the first phase, we train our end-to-end policy network by imitating from a tradition pipeline. In the second phase, we fine-tune our policy network using reinforcement learning to improve the network performance. The framework of our systems is shown as follows.
Fine-tuning network using reinforcement-learning:
Our realworld experiments:
Author: Jiarong Lin
Follow TensorFLow Installation and Pytorch Installation
Following the tutorial of Microsoft airsim, kindly setup your environment.
Our reinforcement-learning is based on OpenAI-baseline platform. However, since we train our network by modifying some of its codes, therefore, our project include the codes of OpenAI-baseline, which is forked form its github.
The following package is needed in this project, you can install the following packages by pip, based on your python's environment settings.
- numpy (for matrix computing)
- openCV2
- transforms3d (for SE3 transformation)
- pickle
Clone DJI Onboard-SDK, switch to branch 3.3
git clone https://github.com/dji-sdk/Onboard-SDK
git checkout 3.3
Install DJI onboard-SDK from the tutorial here.
- Comparison of the trajectory generated from the traditional method and from network.
cd python_scripts/test
python net_vs_tr_and_pl.py
- Test loading policy network
cd python_scripts/test
python test_policy_net.py
- Imitation learning of motion-planning.
- Imitation learning of SE3 geometry controller.
- Environment setup
- Reinforcement-learning
Thanks for Luqi.Wang and Fei.Gao, without their contributions, our works can’t be finished as we expected.
The source code is released under GPLv2 license.
Since I have transferred from the Hong Kong University of Science and Technology (HKUST) to the University of Hong Kong (HKU), and our new lab is under construction, therefore this project is paused for several months. Some of the codes in this project might not be well constructed or well testing. However, we insist on open our code for sharing our discovery, we hope some of our current work can help you. Thank you~