This is the implementation of our RA-L paper arXiv, training and evaluation on AI2-THOR.
- The environment: Cuda 10.0, Python 3.6.4, PyTorch 1.0.1
- Please install AI2-THOR Version 2.1.0 and modify the "controller.py" of AI2-THOR as FILE.
- Our trained model can be downloaded from HERE. If you plan to train your own navigation model from scratch, some suggestions are provided:
- Run the training model: python3 network.py
- Please first run on 20 Kitchen scenes until it converges. Then increase the trining scenes to 40 (20 for kitchen, 20 for living room.) Finally, you can run on 80 training scenes.
- To evaluate our model, please run "python3 eve1_checkpoint.py"
- The files in './test/evaluation1' are for cross-scene evaluation; The files in './test/evaluation2' are for cross-target evaluation.
- The video can be downloaded from HERE.
- Cross-scene generalization on AI2-THOR
- Transfer to the real world
* The navigation system takes as input data from four real-time camera sensors and a target image at each time step, to predict the optimal discrete navigation action.
* The action command is converted to the wheel velocity and passed to the robot.
* For example, the move right action is converted to rotate right at 45^\circ/s for 2s, move forward at 0.25m/s for 2s, and rotate left at 45^\circ/s for 2s.
* These commands are published with a frequency of 5Hz. It is complex due to the movementdirection restrictions of the TurtleBot.
To ask questions or report issues please open an issue on the issues tracker.
If you use this work in your research, please cite the paper:
@article{rlnav2020,
author = {Qiaoyun Wu and
Kai Xu and
Jun Wang and
Mingliang Xu and
Dinesh Manocha},
title = {Reinforcement Learning based Visual Navigation with Information-Theoretic
Regularization},
year = {2019},
url = {http://arxiv.org/abs/1912.04078},
archivePrefix = {arXiv},
eprint = {1912.04078},
}