Gregory Kahn, Pieter Abbeel, Sergey Levine

Website link

• Ubuntu 16.04 (although other versions may work)
• miniconda

Clone this repository.

Setup the anaconda environment:

conda env create -f install/environment.yml

Add the following to your ~/.bashrc file (replacing <SIDEWALK_PATH> appropriately):

activate_sidewalk () {
    export PYTHONPATH=<SIDEWALK_PATH>/src:$PYTHONPATH
    conda activate sidewalk
}

When you want to run the code, make sure to first run activate_sidewalk in your terminal.

You can walk through the data by running the following command:

python scripts/hdf5_visualizer.py -folders experiments/hdf5s

To step through the trajectories, use the keyboard:

q: quit
e: next timestep
w: prev timestep
d: next file
s: prev file
c: next end of file
x: prev end of file
1-9: skip forward by that many timesteps

Create the experiments folder in the root directory

mkdir experiments

Download the contents of this folder and unzip the contents.

Run the evaluation script for our method:

python scripts/eval.py configs/ours.py --model experiments/ours/ckpts/ckpt-7

You can step through the trajectories using the same keyboard commands used for the data visualization.

You can also run the behavioral cloning baseline:

python scripts/eval.py configs/bc.py --model experiments/bc/ckpts/ckpt-1

If you would like to train the models yourself, first delete experiments/ours and experiments/bc. Then train our method:

python scripts/train.py configs/ours.py

and then train the behavioral cloning method:

python scripts/train.py configs/bc.py