This repository accompanies the following paper:
COG: Connecting New Skills to Past Experience with Offline Reinforcement Learning
Avi Singh, Albert Yu, Jonathan Yang, Jesse Zhang, Aviral Kumar, Sergey Levine
Conference on Robot Learning, 2020
Website | Arxiv | Video
Open drawer, take object out | Close top drawer, take object out | Remove obstacle, take object out |
---|---|---|
In this paper, we propose an approach to incorporate a large amount of prior data, either from previously solved tasks or from unsupervised or undirected environment interaction, to extend and generalize learned behavior. This prior data is not specific to any one task, and can be used to extend a variety of downstream skills. We train our policies in an end-to-end fashion, mapping high-dimensional image observations to low-level robot control commands.
This code is based on the original CQL implementation.
By default, all logs will be stored in cog/data/
. If you would like to save to
a different directory, update CUSTOM_LOG_DIR
in the relevant launch script.
An example command for using our method in offline mode:
python examples/cog.py --env=Widow250PickTray-v0 --max-path-length=40 --prior-buffer=pickplace_prior.npy --task-buffer=pickplace_task.npy
An example command for online finetuning from a saved checkpoint:
python examples/cog_finetune.py --checkpoint-dir=LOG_DIR --online-data-only --checkpoint-epoch=1000
An example command for running the behavior cloning baseline:
python examples/chaining_bc.py --env=Widow250PickTray-v0 --max-path-length=40 --prior-buffer=pickplace_prior.npy --task-buffer=pickplace_task.npy
The datasets mentioned above can be downloaded from this Google drive link.
Here are the exact commands to reproduce all results for our method in the paper:
python cog.py --env=Widow250DoubleDrawerOpenGraspNeutral-v0 --max-path-length=50 --prior-buffer=closed_drawer_prior.npy --task-buffer=drawer_task.npy
python cog.py --env=Widow250DoubleDrawerCloseOpenGraspNeutral-v0 --max-path-length=80 --prior-buffer=blocked_drawer_1_prior.npy --task-buffer=drawer_task.npy
python cog.py --env=Widow250DoubleDrawerPickPlaceOpenGraspNeutral-v0 --max-path-length=80 --prior-buffer=blocked_drawer_2_prior.npy --task-buffer=drawer_task.npy
Replacing cog.py
with chaining_bc.py
will allow reproducing experiments
for the BC baseline.
Our code is based on CQL, which is in turn based on rlkit. The setup instructions are similar to rlkit, but we repeat them here for convenience:
conda env create -f environment/linux-gpu-env.yml
source activate cql-env
pip install -e .
After the above, please install roboverse and its dependencies in the same conda env.
The datasets used in this project can be downloaded using this Google drive link.
If you would like to download the dataset on a remote machine via the command line, consider using gdown.
- There is an OpenCV/ffmpeg warning when saving videos. It does not effect any of RL functionality, and the saved videos can still be viewed. However, we will fix this soon.
- There is a torch checkpoint loading warning for online fine-tuning experiments. This also does not effect functionality as long as the correct checkpoints are being loaded. We plan to look into this as well.
- Remove hard-coded image key in line 114 in
rlkit/samplers/rollout_functions.py
- Fix OpenCV/ffmpeg video saving warning - just need to switch to a difference codec.
- Fix torch checkpoint loading warning -- maybe explicitly saving all the networks and optimizer instead of the trainer object can help here.
- Clean up
rlkit/data_management/load_buffer.py
-
ObsDictReplayBuffer
andObsDictRelabelingBuffer
share a lot of code, but are currently implemented independently. - Hard-coded
.cuda()
in line 234 inrlkit/torch/sac/cql.py
- Perhaps the
if
statement (line 437) for image format conversion is a bit too hacky inObsDictReplayBuffer
.