Replacing Rewards with Examples: Example-Based Policy Search via Recursive Classification

This is the pytorch implementation of "Replacing Rewards with Examples: Example-Based Policy Search via Recursive Classification". The original tensorflow version could be found here.

Currently only supports the training of env door-human-v0. The support of the training of other environments will come out subsequently.

Requirements

• python 3.7
• register wandb account
• mujoco
• other packages can be found in requirements.txt

Setup the environment

pip install -e .
git clone https://github.com/rail-berkeley/d4rl.git
cd d4rl
pip install -e .

TODO List

• support the training of other envs in the metaworld.

Reproduce experiments

All the arguments can be found in argments.py.

python trainer.py

Results

Door-human-v0