Skip to content
Code for "Benchmarking Classic and Learned Navigation in Complex 3D Environments" paper
Python Jupyter Notebook Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data Initial commit Jan 25, 2019
examples
navigation Initial commit Jan 25, 2019
README.md Added links to the paper, video and website Jan 31, 2019
benchmark_all_handcrafted_agents.py Initial commit Jan 25, 2019
benchmark_handcrafted_agent.py Initial commit Jan 25, 2019
install_minos_and_deps.sh Added nvidia version detection for headless-gl Jan 31, 2019

README.md

Code for the paper "Benchmarking Classic and Learned Navigation in Complex 3D Environments"

Project website: https://sites.google.com/view/classic-vs-learned-navigation

Video: https://www.youtube.com/watch?v=b1S5ZbOAchc

Paper: https://arxiv.org/abs/1901.10915

If you use this code or the provided environments in your research, please cite the following:

@ARTICLE{Navigation2019,
       author = {{Mishkin}, Dmytro and {Dosovitskiy}, Alexey and {Koltun}, Vladlen},
        title = "{Benchmarking Classic and Learned Navigation in Complex 3D Environments}",
         year = 2019,
        month = Jan,
archivePrefix = {arXiv},
       eprint = {1901.10915},
}

Dependencies:

  • minos
  • numpy
  • pytorch
  • ORBSLAM2

Tested with:

  • Ubuntu 16.04
  • python 3.6
  • pytorch 0.4, 1.0

Benchmark

You may want to comment/uncomment needed agents and/or environments if need to reproduce only part of them. Agents contain random parts: RANSAC in ORBSLAM and 10% random actions in all agents. Nevertheless, results should be the same if run on the same PC. From machine to machine, results may differ (slightly)

You may also want to turn on recording of videos (RGB, depth, GT map, map, beliefs) by setting VIDEO=True in benchmark_all_handcrafted_agents.py.

Simple example of working with agents is shown in example.

Training

Training and pre-trained weights for the learned agents are coming soon.

You can’t perform that action at this time.