This is the code for our paper "Mutual Information State Intrinsic Control".
The code was developed by Rui Zhao during a research internship at Horizon Robotics Inc. Cupertino, CA, USA.
The paper is accepted by International Conference on Learning Representations (ICLR) 2021 as Spotlight.
The paper is available on ICLR openreview: https://openreview.net/forum?id=OthEq8I5v1.
Our code is based on OpenAI Baselines (link: https://github.com/openai/baselines).
The code requires python3 (>=3.5) with the development headers. You'll also need system packages CMake, OpenMPI and zlib. Those can be installed as follows
sudo apt-get update && sudo apt-get install cmake libopenmpi-dev python3-dev zlib1g-devTo run the code, you need to install OpenAI Gym (link: https://github.com/openai/gym).
We use the robotics environment in OpenAI Gym, which needs the MuJoCu physics engine (link: http://www.mujoco.org/).
The experiments were carried out on a 16-CPUs server.
We use 16 CPUs for training.
If you are running the experiments on a laptop, please configure a smaller number of CPUs.
Note that, with less CPUs, the performance will be affected.
After the installation of dependencies, you can start to reproduce the experimental results by running the following commands.
The following command is for training an agent using MUSIC without any rewards or supervision.
python baselines/her/experiment/train.py --env_name FetchPickAndPlace-v1 --n_epochs 50 --num_cpu 16 --logging True --seed 0 --note SAC+MUSIC
To test the learned policies, you can run the command:
python baselines/her/experiment/play.py /path/to/an/experiment/policy_latest.pkl --note SAC+MUSIC
The rendered video is saved alongside the policy file.
You can also extract the weight from the saved TensorFlow graph .pkl file. To do this, you need to put the path of the .pkl file into the config file, for example, into "params/SAC+MUSIC.json". Afterwards, you can use the following command to convert the .pkl file:
python baselines/her/experiment/save_weight.py --env_name FetchPickAndPlace-v1 --note SAC+MUSIC-r --seed 0
After converting, you can use the pre-trained MI discriminator to accelerate learning:
python baselines/her/experiment/train.py --env_name FetchPickAndPlace-v1 --n_epochs 50 --num_cpu 16 --logging True --seed 0 --note SAC+MUSIC-r
Citation of the paper:
@inproceedings{zhao2021mutual,
title={Mutual Information State Intrinsic Control},
author={Zhao, Rui and Gao, Yang and Abbeel, Pieter and Tresp, Volker and Xu, Wei},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=OthEq8I5v1}
}
MIT