This repository accompanies the following paper:

A Workflow for Offline Model-Free Robotic RL
Aviral Kumar*, Anikait Singh*, Stephen Tian, Chelsea Finn, Sergey Levine
Conference on Robot Learning, 2021
Website | Arxiv | Video

Offline reinforcement learning (RL) enables learning control policies by utilizing only prior experience, without any online interaction. This can allow robots to acquire generalizable skills from large and diverse datasets, without any costly or unsafe online data collection. Despite recent algorithmic advances in offline RL, applying these methods to real-world problems has proven challenging. Although offline RL methods can learn from prior data, there is no clear and well-understood process for making various design choices, from model architecture to algorithm hyperparameters, without actually evaluating the learned policies online. In this paper, our aim is to develop a practical workflow for using offline RL analogous to the relatively well-understood workflows for supervised learning problems. To this end, we devise a set of metrics and conditions that can be tracked over the course of offline training, and can inform the practitioner about how the algorithm and model architecture should be adjusted to improve final performance. Our workflow is derived from a conceptual understanding of the behavior of conservative offline RL algorithms and cross-validation in supervised learning. We demonstrate the efficacy of this workflow in producing effective policies without any online tuning, both in several simulated robotic learning scenarios and for three tasks on two distinct real robots, focusing on learning manipulation skills with raw image observations with sparse binary rewards.

This code is based on the COG implementation.

In this paper, two offline RL algorithms were tested: CQL and BRAC. We utilized these algorithms as baseline on which we applied our workflow. Below are some sample commands to replicate baseline perfromance.

# CQL baseline
python examples/cql_workflow.py --env=Widow250DoubleDrawerOpenGraspNeutral-v0 --max-path-length=50 --buffer 0 --num_traj 100
python examples/cql_workflow.py --env=Widow250DoubleDrawerCloseOpenGraspNeutral-v0 --max-path-length=80 --buffer 1 --num_traj 100
python examples/cql_workflow.py --env=Widow250DoubleDrawerPickPlaceOpenGraspNeutral-v0 --max-path-length=80 --buffer 2 --num_traj 100

# BRAC baseline
python examples/brac_workflow.py --env=Widow250DoubleDrawerOpenGraspNeutral-v0 --max-path-length=50 --buffer 0 --num_traj 100
python examples/brac_workflow.py --env=Widow250DoubleDrawerCloseOpenGraspNeutral-v0 --max-path-length=80 --buffer 1 --num_traj 100
python examples/brac_workflow.py --env=Widow250DoubleDrawerPickPlaceOpenGraspNeutral-v0 --max-path-length=80 --buffer 2 --num_traj 100

Parts of the workflow can be applied using their respective flags. For example, to run the CQL algorithm with the VIB architecture utilized in the overfitting regime, you can run the following command:

python examples/cql_workflow.py --env=Widow250DoubleDrawerPickPlaceOpenGraspNeutral-v0 --max-path-length=80 --buffer 2 --num_traj 100 --bottleneck --bottlneck_const=1e-3

Here is a list of other flags that are available for the CQL and BRAC algorithms. I have listed some relevant flags and their usage.

• General Hyperparameters
  • '--env': Environment to train on.
  • '--buffer': Buffer to use.
  • '--max-path-length': Maximum length of a single rollout.
• CQL algorithm
  • '--min_q_weight': CQL alpha parameter. Controls conservativeness of the algorithm.
• Brac algorithm
  • '--beta': Value of beta in BRAC
  • '--behavior_path': Path to behavior policy
• Workflow Corrections
  • '--bottleneck': Use the VIB architecture. (Overfitting Correction)
  • '--bottleneck_const': Regularization constant for the VIB.
  • '--num_traj': Number of trajectories to collect.
  • '--dr3_feat': Use DR3 regularization. (Underfitting Correction)
  • '--dr3_weight': Weight for DR3 regularization.
  • '--regularization': Enable L1/L2 regularization. (Overfitting Correction)
  • '--regularization_type': L1/L2 regularization type.
  • '--dropout': Enable dropout regularization. (Overfitting Correction)
  • '--resnet_enc': Use encoder with residual layers (Underfitting Correction)

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 create --name workflow python=3.8
conda activate workflow
pip install -r requirements.txt

After the above, please install the following fork of roboverse by running:

python setup.py develop

The datasets used in this project can be downloaded using this Google drive link. (to be updated)

If you would like to download the dataset on a remote machine via the command line, consider using gdown.