Enhanced Meta Reinforcement Learning using Demonstrations in Sparse Reward Environments

Code for Enhanced Meta Reinforcement Learning using Demonstrations in Sparse Reward Environments (NeurIPS 2022)

This code is based on a public meta-rl github repository learn2learn

Video of real world demonstrations on a TurtleBot

To run expirements you will need the the following packages:

• cherry-rl 0.1.4
• tensorboard
• learn2learn
• mujoco-py 2.0.2.13
• torch 1.10.0
• gym 0.21.1

To run the experiments, simply execute the following commands,

Particle2D Navigation:

python EMRLD_PN.py    --exp-num i 
python EMRLD-WS_PN.py --exp-num i

Two Wheeled Locomotion:

python EMRLD_TW.py    --exp-num i   
python EMRLD-WS_TW.py --exp-num i

HalfCheetah Forward-Backward:

python EMRLD_HC.py    --exp-num i   
python EMRLD-WS_HC.py --exp-num i

Where i = 1 is for Optimal demonstration data, and i = 2 is for sub-optimal demonstration data