Learning to navigate without a map

The goal of the project is to determine how well deep learning is suited for planning under incomplete information.

Instructions

Requirements

• Python 2.7 (In principle, this project can be run under Python 3, we didn't test it however)
• Keras 2
• Some specific packages in requirements.txt

D-star package compilation

Please follow the instruction in Dstar implementation

How to use this package

1. Clone this package:

$ git clone https://github.com/ToniRV/Learning-to-navigate-without-a-map

2. Check you have the project resource folder at

$HOME/.rlvision

Note that this folder will be automatically created at the first time that you run the package, you can get the correct resource folder by

$ python ./rlvision/__init__.py

3. Copy data to the data sub-folder in $HOME/.rlvision/

Download the data from here.

Uncompress the folder and place all files in data sub-folder

Run Experiments

1. Run VIN experiments via

$ make vin-exp-po-8   # for 8x8 grid world
$ make vin-exp-po-16  # for 16x16 grid world
$ make vin-exp-po-28  # for 28x28 grid world

2. Run PG experiments via

$ make pg-16-exp    # policy gradients
$ make ddpg-16-exp  # deep determinstic policy gradients

3. Run DQN experiments via

$ make dqn-8-exp   # for 8x8 grid world
$ make dqn-16-exp  # for 16x16 grid world
$ make dqn-28-exp  # for 28x28 grid world

4. Run D-star experiments via

$ make dstar-8-exp   # for 8x8 grid world
$ make dstar-16 exp  # for 16x16 grid world
$ make dstar-28 exp  # for 28x28 grid world

Contacts

Yuhuang Hu, Shu Liu, Antoni Rosiñol Vidal, Yang Yu
Email: {hyh, liush, antonir, yuya}@student.ethz.ch