Continual World
Continual World is a benchmark for continual reinforcement learning. It contains realistic robotic tasks which come from MetaWorld.
The core of our benchmark is CW20 sequence, in which 20 tasks are run, each with budget of 1M steps.
We provide the complete source code for the benchmark together with the tested algorithms implementations and code for producing result tables and plots.
See also the paper and the website.
Installation
You can either install directly in Python environment (like virtualenv or conda), or build containers -- Docker or Singularity.
Standard installation (directly in environment)
First, you'll need MuJoCo simulator. Please follow the instructions from mujoco_py
package. As MuJoCo has been made freely available, you can obtain a free license here.
Next, go to the main directory of this repo and run
pip install .
Alternatively, if you want to install in editable mode, run
pip install -e .
Docker image
-
To build the image with
continualworld
package installed inside, rundocker build . -f assets/Dockerfile -t continualworld
-
To build the image WITHOUT the
continualworld
package but with all the dependencies installed, rundocker build . -f assets/Dockerfile -t continualworld --build-arg INSTALL_CW_PACKAGE=false
When the image is ready, you can run
docker run -it continualworld bash
to get inside the image.
Singularity image
-
To build the image with
continualworld
package installed inside, runsingularity build continualworld.sif assets/singularity.def
-
To build the image WITHOUT the
continualworld
package but with all the dependencies installed, runsingularity build continualworld.sif assets/singularity_only_deps.def
When the image is ready, you can run
singularity shell continualworld.sif
to get inside the image.
Running
You can run single task, continual learning or multi-task learning experiments with run_single.py
, run_cl.py
, run_mt.py
scripts, respectively.
To see available script arguments, run with --help
option, e.g.
python3 run_single.py --help
Examples
Below are given example commands that will run experiments with a very limited scale.
Single task
python3 run_single.py --seed 0 --steps 2e3 --log_every 250 --task hammer-v1 --logger_output tsv tensorboard
Continual learning
python3 run_cl.py --seed 0 --steps_per_task 2e3 --log_every 250 --tasks CW20 --cl_method ewc --cl_reg_coef 1e4 --logger_output tsv tensorboard
Multi-task learning
python3 run_mt.py --seed 0 --steps_per_task 2e3 --log_every 250 --tasks CW10 --use_popart True --logger_output tsv tensorboard
Reproducing the results from the paper
Commands to run experiments that reproduce main results from the paper can be found in examples/paper_cl_experiments.sh
, examples/paper_mt_experiments.sh
and examples/paper_single_experiments.sh
. Because of number of different runs that these files contain, it is infeasible to just run it in sequential manner. We hope though that these files will be helpful because they precisely specify what needs to be run.
After the logs from runs are gathered, you can produce tables and plots - see the section below.
Producing result tables and plots
After you've run experiments and you have saved logs, you can run the script to produce result tables and plots:
python produce_results.py --cl_logs examples/logs/cl --mtl_logs examples/logs/mtl --baseline_logs examples/logs/baseline
In this command, respective arguments should be replaced for paths to directories containing logs from continual learning experiments, multi-task experiments and baseline (single-task) experiments. Each of these should be a directory inside which there are multiple experiments, for different methods and/or seeds. You can see the directory structure in the example logs included in the command above.
Results will be produced and saved on default to the results
directory.
Alternatively, check out nb_produce_results.ipynb
notebook to see plots and tables in the notebook.
Download our saved logs and produce results
You can download logs of experiments to reproduce paper's results from here. Then unzip the file and run
python produce_results.py --cl_logs saved_logs/cl --mtl_logs saved_logs/mt --baseline_logs saved_logs/single
to produce tables and plots.
As a result, a csv file with results will be produced, as well as the plots, like this one (and more!):
Full output can be found here.
Acknowledgements
Continual World heavily relies on MetaWorld.
The implementation of SAC used in our code comes from Spinning Up in Deep RL.
Our research was supported by the PLGrid infrastructure.
Our experiments were managed using Neptune.