Target-driven Visual Navigation Model using Deep Reinforcement Learning

THOR scene samples

Introduction

This repocitory provides a Tensorflow implementation of the deep siamese actor-critic model for indoor scene navigation introduced in the following paper:

Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning
Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi
ICRA 2017, Singapore

Setup

This code is implemented in Tensorflow API r1.0. You can follow the online instructions to install Tensorflow 1.0. Other dependencies (h5py, numpy, scikit-image, pyglet) can be install by pip: pip install -r requirements.txt. This code has been tested with Python 2.7 and 3.5.

Scenes

To facilitate training, we provide hdf5 dumps of the simulated scenes. Each dump contains the agent's first-person observations sampled from a discrete grid in four cardinal directions. To be more specific, each dump stores the following information row by row:

Before running the code, please download the scene dumps using the following script:

./data/download_scene_dumps.sh

We are currently releasing one scene from each of the four scene categories, bathroom, bedroom, kitchen, and living room. Please contact me for information about additional scenes. A keyboard_agent.py script is provided. This script allows you to load a scene dump and use the arrow keys to navigate a scene. To run the script, here is an example command:

# make sure the scene dump is in the data folder, e.g., ./data/bedroom_04.h5
python keyboard_agent.py --scene_dump ./data/bedroom_04.h5

These scene dumps enable us to train a (discrete) navigation agent without running the simulator during training or extracting ResNet features. Thus, it greatly improves training efficiency. The training code runs comfortably on CPUs (of my Macbook Pro). Due to legal concerns, our THOR simulator will be released later.

Training and Evaluation

The parameters for training and evaluation are defined in constants.py. The most important parameter is TASK_LIST, which is a dictionary that defines the scenes and targets to be trained and evaluated on. The keys of the dictionary are scene names, and the values are a list of location ids in the scene dumps, i.e., navigation targets. We use a type of asynchronous advantage actor-critic model, similar to A3C, where each thread trains for one target of one scene. Therefore, make sure the number of training threads PARALLEL_SIZE is at least the same as the total number of targets. You can use more threads to further parallelize training. For instance, when using 8 threads to train 4 targets, 2 threads will be allocated to train each target.

The model checkpoints are stored to CHECKPOINT_DIR, and Tensorboard logs are written in LOG_FILE. To train a target-driven navigation model, run the following script:

# train a model for targets defined in TASK_LIST
python train.py

For evaluation, we run 100 episodes for each target and report the mean/stddev length of the navigation trajectories. To evaluate a model checkpoint in CHECKPOINT_DIR, run the following script:

# evaluate a checkpoint on targets defined in TASK_LIST
python evaluate.py

Acknowledgements

I would like to acknowledge the following references that have offered great help for me to implement the model.

Citation

Please cite our ICRA'17 paper if you find this code useful for your research.

@InProceedings{zhu2017icra,
  title = {{Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning}},
  author = {Yuke Zhu and Roozbeh Mottaghi and Eric Kolve and Joseph J. Lim and Abhinav Gupta and Li Fei-Fei and Ali Farhadi},
  booktitle = {{IEEE International Conference on Robotics and Automation}},
  year = 2017,
}

License

MIT