This repository contains the code release for the paper Self-Supervised Visual Reinforcement Learning with Object-Centric Representations by Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius, published as a spotlight at ICLR 2021. Please use the provided citation when making use of our code or ideas.
The code is an adaptation of RLkit for RL code and multiworld for environments, both under MIT license.
Autonomous agents need large repertoires of skills to act reasonably on new tasks that they have not seen before. However, acquiring these skills using only a stream of high-dimensional, unstructured, and unlabeled observations is a tricky challenge for any autonomous agent. Previous methods have used variational autoencoders to encode a scene into a low-dimensional vector that can be used as a goal for an agent to discover new skills. Nevertheless, in compositional/multi-object environments it is difficult to disentangle all the factors of variation into such a fixed-length representation of the whole scene. We propose to use object-centric representations as a modular and structured observation space, which is learned with a compositional generative world model. We show that the structure in the representations in combination with goal-conditioned attention policies helps the autonomous agent to discover and learn useful skills. These skills can be further combined to address compositional tasks like the manipulation of several different objects.
- Install poetry and make sure that correct version of python is installed.
- Install MuJoCo
- Create and activate a new poetry environment with
poetry installandpoetry shellthat you run from repo folder.
To test that everything installed correctly you can run tests for SMORL:
python -W ignore ./rlkit/settings/smorl/smoke_test.py
and for baselines:
python -W ignore ./rlkit/settings/baselines/sac/smoke_test.py
python -W ignore ./rlkit/settings/baselines/rig/smoke_test.py
python -W ignore ./rlkit/settings/baselines/skewfit/smoke_test.py
All the hyperparameters files are stored in ./rlkit/settings folder.
To run SMORL training, just execute the corresponding exp_name.py file. For example, for SMORL training on GT representations in 3 objects Rearrange environment run
python ./rlkit/settings/smorl/gt_rearrange_3_object.py
For Visual Environments, we have provided a trained SCALOR model. If you want you can train SCALOR model from scratch by running
python ./rlkit/settings/scalor_training/exp_name.py
After SCALOR training, you can use the saved checkpoint path in the corresponding SMORL training setting by
variant["scalor_path"]="path/to/checkpoint"
You can run a visualization of the SMORL training results by
python -m viskit.frontend ./rlkit/data/exp_name/
If you want additionally visualize SCALOR training please run tensorboard with SCALOR training results dir.
Finally, you can visualize learned policy by running
python rlkit/scripts/run_goal_conditioned_policy.py ./rlkit/data/exp_name/params.pkl
Please use the following bibtex entry to cite us:
@inproceedings{Zadaianchuk2021SelfSupervisedVRL,
title={Self-supervised Visual Reinforcement Learning with Object-centric Representations},
author={Andrii Zadaianchuk and Maximilian Seitzer and Georg Martius},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=xppLmXCbOw1}
}
We used RLkit, multiworld and viskit by Vitchyr Pong for RL infrastructure and SAC+HER training implementation as well as environments implementations. Also, the SCALOR implementation is an adapted version of the official SCALOR implementation by Jindong Jiang.