Language-conditioned Meta-Reinforcement Learning for Multi Manipulation Tasks

This repository contains the official code of master's thesis work "Language-conditioned Meta-Reinforcement Learning for Multi Manipulation Tasks".

In order to repeat the experiments, run the script as:

python main_async.py

This will execute the experiment with a random seed, selected by the rlpyt. In order to replicate the results presented in the work,

python main_async.py --seed 0
python main_async.py --seed 1
python main_async.py --seed 2

Results of the experiments are then saved under the directory experiments in the following structure:

run_[DAY][MONTH][YEAR]-[HOUR][MINUTE][SECOND]

Creating figures

To create the figures in the thesis work, you can run the make_figures.py script in the figures directory as

python make_figures.py

This will create all the figures and store them inside the figures directory.

Generating Contextual Embeddings

To generate contextual embeddings for the environment name, run context_embeddings.py in environment directory as

python context_embeddings.py

This will save the embeddings as a dict structure in which keys are the environment names and the values are the context embeddings, and it is pickled into the file context_embeddings_roberta.pkl.