DC2G: Deep Cost-to-Go Planning Algorithm (IROS '19)

Planning Beyond the Sensing Horizon Using a Learned Context

Michael Everett, Justin Miller, Jonathan P. How

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019

Paper: https://arxiv.org/abs/1908.09171

Video: https://youtu.be/yVlnbqEFct0

Note: This repo contains the following code that was approved for release:

• Bing Maps Driveway Dataset (~80 houses)
• Pre-trained cost-to-go estimation network
• Gridworld evaluation environment (built on gym-minigrid)
• Jupyter Notebook to explain code

Instructions

Install

git clone --recursive <this repo>
cd dc2g
./install.sh

See an Interactive Example in Jupyter Notebook

Either see dc2g.ipynb or run this script to open an interactive notebook in the virtualenv with dependencies installed already:

./notebook.sh

Run Example (streamlined version of Jupyter notebook)

This will initialize the environment and run an episode with a DC2G planner:

./example.sh

Repo Structure/Organization

Bing Maps Driveway Dataset:

• data/datasets/driveways_bing_iros19 contains .pngs of the 80 houses split into train/val/test, in 3 formats:
  • raw: satellite image
  • full_semantic: same size as satellite image, but each pixel colored by semantic class
  • full_c2g: same size as satellite image, but each pixel colored by grayscale intensity corresponding to cost-to-go to door pixels. Non-traversable terrain (grass) is colored red.
• data/scripts: contains various python scripts to compute cost-to-go, apply masks, etc. to generate the training datasets

Pre-trained cost-to-go estimation network:

• data/trained_networks/driveways_bing_iros19_full_test_works holds the checkpoint, export.index, and export.meta files needed to load the network in Python. We also include options.json to see hyperparameters used to train to generate this network.

DC2G Planner Code:

• dc2g/planners contains several planners, all of which inherit from Planner.py.

Environment Evaluation code:

• dc2g/run_episode.py instantiates the environment and a planner and runs one episode
• dc2g/run_experiment.py instantiates the environment and runs many episodes with many planners to compute statistics

Environment code:

• gym-minigrid follows the OpenAI Gym API (env.step,env.reset, etc.) and the particular python script of interest here is gym-minigrid/gym_minigrid/envs/slam.py, which inherits from gym-minigrid/gym_minigrid/minigrid.py and gives extra capabilities to build up a map of the env over time.

Jupyter notebook:

• dc2g.ipynb can be viewed in a browser
• notebook.sh opens a local Jupyter notebook instance so you can interactively explore dc2g.ipynb in the virtualenv with proper dependencies installed

TODOs

• Get OraclePlanner working again
• Get FrontierPlanner working again
• Get DC2GPlanner working again
• Get DC2GRescalePlanner working again
• Create Jupyter notebook to explain code
• Confirm run_episode.py works
• Confirm run_experiment.py works
• Add instructions for generating masked semantic/c2g images
• Add scripts to generate plots from paper
• Pull our custom slam.py code out of gym-minigrid -- see driveway_env.py which subclasses MiniGridEnv from gym-minigrid-mit (python 2.7 compatible version of gym-minigrid)

If you find this code useful, please consider citing our paper:

@inproceedings{Everett19_IROS,
	Address = {Macau, China},
	Author = {Michael Everett and Justin Miller and Jonathan P. How},
	Booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
	Date-Modified = {2019-06-22 06:18:08 -0400},
	Month = {November},
	Title = {Planning Beyond The Sensing Horizon Using a Learned Context},
	Year = {2019}
}