Don't Change the Algorithm, Change the Data:

Exploratory Data for Offline Reinforcement Learning

Denis Yarats*, David Brandfonbrener*, Hao Liu, Michael Laskin,

Pieter Abbeel, Alessandro Lazaric, Lerrel Pinto

Paper: https://arxiv.org/abs/2201.13425

Code and Data: https://github.com/denisyarats/exorl

Abstract

Recent progress in deep learning has relied on access to large and diverse datasets. Such data-driven progress has been less evident in offline reinforcement learning (RL), because offline RL data is usually collected to optimize specific target tasks limiting the data's diversity. In this work, we propose Exploratory data for Offline RL (ExORL), a data-centric approach to offline RL. ExORL first generates data with unsupervised reward-free exploration, then relabels this data with a downstream reward before training a policy with offline RL. We find that exploratory data allows vanilla off-policy RL algorithms, without any offline-specific modifications, to outperform or match state-of-the-art offline RL algorithms on downstream tasks. Our findings suggest that data generation is as important as algorithmic advances for offline RL and hence requires careful consideration from the community.

ExORL framework

The ExORL framework has three phases, illustrated to the left: data collection with unsupervised RL, data relabeling with the downstream reward, and learning on the labeled data with offline RL

Some results

To the left we show results of Offline RL evaluation using vanilla TD3 on supervised data (the replay buffer of a TD3 agent trained on Walker Stand) versus unsupervised exploratory data (collected reward-free on Walker with Proto). Unsupervised data demonstrates an average performance improvement of 106% across the four evaluation tasks.

Below we show an illustrative example of the ExORL framework on a pointmass maze environment. We run nine different exploratory data collection algorithms and five different offline RL algorithms on four downstream tasks and plot the results.

See the paper for full results.

Using the data

Our github repo (https://github.com/denisyarats/exorl) contains full details about how to download and re-label the exploratory data that we generated.

We hope that this data can provide a distinct challenge to offline RL algorithms compared to prior work that uses supervised data generation where the policy that generates the data is attempting to optimize the same reward function as the offline RL algorithm.