CMTA

(NIPS23)Contrastive Modules with Temporal Attention for Multi-Task Reinforcement Learning

This repository is based on MTRL. Please refer to the documentation of MTRL for any installation issues or guidance.

Before running our code, you first need to install MetaWorld successfully. We use the af8417bfc82a3e249b4b02156518d775f29eb289 commit for the MetaWorld environments for our experiments.

Notice

We use experiment.random_pos to choose Fixed environment or Mixed environment, and add corresponding code in src/mtenv, so it doesn't need to git clone submodule mtenv anymore.

Getting Started

We should install the local mtenv lib at first:

cd src/mtenv
pip install -e .

Then you can use the following instructions to run CMTA:

cd scripts
bash CMTA.sh $seed$

\$ seed \$ can be 1,10,42,...