Hierarchical Reinforcement Learning based on Planning Operators

Introduction

Paper website: https://arxiv.org/abs/2309.14237

Authors: Jing Zhang, Karinne Ramirez-Amaro

Framework

Setup

[email protected]:jingzhang00/RL_operator.git

pip install -r requirements.txt

Training

cd scripts/experiments
bash main.bash 42 cuda:0 stack_2

Independent Policy Evaluation

without render

cd scripts/evaluation
bash visualize_model.bash 37 "trained_model/" "state_dict.pt" "sacx_experiment_setting.pkl" 50 5 false

render

bash visualize_model.bash 37 "trained_model/" "state_dict.pt" "sacx_experiment_setting.pkl" 50 5 true

for other policies, change the last number "5", (open:0, close:1, reach:2, lift:3, move:4, stack:5)

Note

1. Current hyperparameters only apply for 2 blocks, more blocks need fine-tuning.
2. As for planning success rate, it was evaluated through training process, check "/home/omen/Downloads/RL_operator/trained_model/tensorboard":

cd trained_model
tensorboard --logdir=tensorboard

then it is shown in "evaluation_info/epoch_success_rate".