April 9, 2020

505 words 3 mins read



Pytorch code for ICLR-20 Paper "Learning to Explore using Active Neural SLAM"

repo name devendrachaplot/Neural-SLAM
repo link https://github.com/devendrachaplot/Neural-SLAM
homepage http://www.cs.cmu.edu/~dchaplot/projects/neural-slam.html
language Python
size (curr.) 6287 kB
stars (curr.) 158
created 2020-04-10
license MIT License

Active Neural SLAM

This is a PyTorch implementation of the ICLR-20 paper:

Learning To Explore Using Active Neural SLAM Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, Ruslan Salakhutdinov Carnegie Mellon University, Facebook AI Research, UIUC

Project Website: https://devendrachaplot.github.io/projects/Neural-SLAM



The Active Neural SLAM model consists of three modules: a Global Policy, a Local Policy and a Neural SLAM Module. As shown below, the Neural-SLAM module predicts a map and agent pose estimate from incoming RGB observations and sensor readings. This map and pose are used by a Global policy to output a long-term goal, which is converted to a short-term goal using an analytic path planner. A Local Policy is trained to navigate to this short-term goal.


Installing Dependencies

We use earlier versions of habitat-sim and habitat-api. The specific commits are mentioned below.

Installing habitat-sim:

git clone https://github.com/facebookresearch/habitat-sim.git
cd habitat-sim; git checkout 9575dcd45fe6f55d2a44043833af08972a7895a9; 
pip install -r requirements.txt; 
python setup.py install --headless
python setup.py install # (for Mac OS)

Installing habitat-api:

git clone https://github.com/facebookresearch/habitat-api.git
cd habitat-api; git checkout b5f2b00a25627ecb52b43b13ea96b05998d9a121; 
pip install -e .

Install pytorch from https://pytorch.org/ according to your system configuration. The code is tested on pytorch v1.2.0. If you are using conda:

conda install pytorch==1.2.0 torchvision cudatoolkit=10.0 -c pytorch #(Linux with GPU)
conda install pytorch==1.2.0 torchvision==0.4.0 -c pytorch #(Mac OS)


Clone the repository and install other requirements:

git clone --recurse-submodules https://github.com/devendrachaplot/Neural-SLAM
cd Neural-SLAM;
pip install -r requirements.txt

The code requires datasets in a data folder in the following format (same as habitat-api):


Please download the data using the instructions here: https://github.com/facebookresearch/habitat-api#data

To verify that dependencies are correctly installed and data is setup correctly, run:

python main.py -n1 --auto_gpu_config 0 --split val



For training the complete Active Neural SLAM model on the Exploration task:

python main.py

Downloading pre-trained models

mkdir pretrained_models
wget -O pretrained_models/model_best.global http://www.cs.cmu.edu/~dchaplot/projects/active_neural_slam/model_best.global
wget -O pretrained_models/model_best.local http://www.cs.cmu.edu/~dchaplot/projects/active_neural_slam/model_best.local
wget -O pretrained_models/model_best.slam http://www.cs.cmu.edu/~dchaplot/projects/active_neural_slam/model_best.slam

For evaluation:

For evaluating the pre-trained models:

python main.py --split val --eval 1 --train_global 0 --train_local 0 --train_slam 0 \
--load_global pretrained_models/model_best.global \
--load_local pretrained_models/model_best.local \
--load_slam pretrained_models/model_best.slam 

For visualizing the agent observations and predicted map and pose, add -v 1 as an argument to the above

For more detailed instructions, see INSTRUCTIONS.

Cite as

Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A. and Salakhutdinov, R., 2020. Learning To Explore Using Active Neural SLAM. In International Conference on Learning Representations (ICLR). (PDF)


  title={Learning To Explore Using Active Neural SLAM},
  author={Chaplot, Devendra Singh and Gandhi, Dhiraj and Gupta, Saurabh and Gupta, Abhinav and Salakhutdinov, Ruslan},
  booktitle={International Conference on Learning Representations (ICLR)},


This repository uses Habitat API (https://github.com/facebookresearch/habitat-api) and parts of the code from the API. The implementation of PPO is borrowed from https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/. We thank Guillaume Lample for discussions and coding during initial stages of this project.

comments powered by Disqus