/Residual-Policy-Learning

Implementation of Residual Policy Learning

Primary LanguagePython

Residual-Policy-Learning

6.881 Robotic Manipulation Course Project, MIT Fall 2020 - Implementation of Residual Policy Learning

Results:

Push Task

Push

Pick And Place Task

PickPlace

Slide Task

Slide

Usage:

First install the requirements using:

pip install -r requirements.txt or pip3 install -r requirements.txt

Make a wandb account:

All the experiments will be logged with a library called "Wandb". So make sure that you first make an account in wandb here and login your terminal using wandb login

Non-MPI version:

python RL/ddpg/ddpg.py --n_cycles=10 --env_name='FetchReach'

MPI version:

mpirun -np 1 python -u RL/ddpg/ddpg_mpi.py --env_name='FetchReach' --n_cycles=10

SAC

python RL/sac/sac.py --env_name="robosuiteNutAssemblyDense

TODO:

  • Improve README with more plots and project information.
  • Add information about all files.
  • Make SAC modular.
  • Add report, video links.
  • If possible add wandb report and link to network weights.