Do deep reinforcement learning algorithms really learn to navigate?
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
BaDhGrICLR2018 @ 6a2e567
deepmind-lab @ b118e72
openai-a3c-impl @ 2449494

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

To install python depedencies for openai

make -f makefiles/

To load the environment paths