Active Exploration for Robotic Manipulation

This repository contains the code for our IROS 2022 paper "Active Exploration for Robotic Manipulation".

Installation

This code requires a python 3.7 or higher installation to run. All requirements can be installed via

pip install -r requirements.txt

Running the experiments presented in the paper

Use the following commands to run the experiments presented in the paper. Each run will create a log dir in results/ which contains checkpoints and the tensorboard log.

Experiments on flat table (Tilted Pushing)

# Mutual Information
python main.py new -e ball_f02h0 --total-episodes 10000 -i mutual_information

# Lautum Information
python main.py new -e ball_f02h0 --total-episodes 10000 -i lautum_information -rs 2e5

# No intrinsic reward
python main.py new -e ball_f02h0 --total-episodes 10000 --no-intrinsic

Experiments on table with holes (Tilted Pushing Maze)

# Mutual Information
python main.py new -e ball_f05h1 --total-episodes 50000 -i mutual_information

# Lautum Information (here we use an adaptive weighting scheme)
python main.py new -e ball_f05h1 --total-episodes 50000 -i lautum_information -rs 2e5 -a max -awrs 1e8

# No intrinsic reward
python main.py new -e ball_f05h1 --total-episodes 50000 --no-intrinsic

If you want to run a different algorithm on these environments, check out https://github.com/TimSchneider42/sisyphus-env, which contains the environments and an example on how to use them.

For a documentation of the options of this software, run

python main.py -h