Implicit mapping by learning to navigate

This repository contains the source code corresponding to the paper: Do Deep Reinforcement Learning algorithms really learn to navigate

Replicating experiments

• Checkout the repository

git clone git@github.com:umrobotslang/does-drl-learn-to-navigate.git

• Install bazel, module, singularity, python packages and compile deepmind lab. To compile inside singularity container read instructions here

cd implicit-mapping
export INSTALL_PREFIX=$(pwd)/build
make build

• To run experiments

cd ./openai-a3c-impl/web/
./run_exp.sh

• To generate plots, edit the ./BaDhGrICLR2018/gennav/exp-results/copy-files.sh to copy json results from experiment location to the exp-results folder.

cd ./BaDhGrICLR2018/gennav/exp-results/
./copy-files.sh

• The provided make file takes care of summarizing the results into csv files and then generating the plots using the csv files.

cd ./BaDhGrICLR2018/gennav/
make exp-results/ntrained.csv ./exp-results/Static_Goal_Random_Spawn_Static_Maze.csv
make images/plot_summary_bar_plots.pdf images/plot_ntrain_summary.pdf

Making deepmind lab

To make deepmind lab

make -f makefiles/deepmind-lab.mk

To install python depedencies for openai

make -f makefiles/openai-a3c.mk

To load the environment paths

source setup.sh