Overview

This is a repository for the paper "Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models".

[Website] [Paper]

<img src="assets/main_idea.png" width="100%" style="max-width: 640px"><br/>

In this paper, we focus on methods that can learn from offline trajectories without reward annotations. We test methods ranging from RL to control, and find that planning with a learned latent dynamics model (PLDM) is a promising approach for this setting when the data is imperfect.

Setting up

Repo Setup

git clone git@github.com:vladisai/PLDM.git

cd PLDM

pip install -r requirements.txt

pip install -e .

Run Experiments

1. Go to pldm_envs/, follow instructions to set up dataset for the environment of your hoice
2. Go to pldm/, follow instruction to run training or evaluation

Datasets

To see the datasets we used to train our models, see folders inside pldm_envs/. The readmes there will guide you on how to download and set up the datasets.