Residual Skill Policies: Learning an Adaptable Skill-based Action Space for Reinforcement Learning for Robotics

[Paper] [Project Page]

Official PyTorch implementation for the publication Residual Skill Policies: Learning an Adaptable Skill-based Action Space for Reinforcement Learning for Robotics (CoRL 2022)

Requirements

• python 3.7+
• mujoco 2.1
• Ubuntu 18.04

Installation Instructions

To install MuJoCo follow the instructions here.

Clone the repository

git clone https://github.com/krishanrana/reskill.git

Ensure conda is installed and configured for your system. Create a conda environment and install all required packages.

cd reskill
conda env create -f environment.yml
conda activate reskill_new
pip install -e .
cd reskill

Data Collection and Training

To collect a dataset using the scripted controllers run the following command:

python data/collect_demos.py --num_trajectories 40000 --subseq_len 10 --task block

There are two sets of tasks block and hook The dataset collected for the block tasks can be used to train a downstream RL agent in the FetchPyramidStack-v0, FetchCleanUp-v0 and FetchSlipperyPush-v0 environments. The dataset collected for the hook task is used to train the downstream RL agent in the FetchComplexHook-v0 environment. We collect the demonstration data for the hook and block based environments in the FetchHook-v0 and FetchPlaceMultiGoal-v0 environments respectively.

You can alternatively download a pre-collected dataset from here and place the unzipped dataset folder in the root of the repository.

To train the skill modules on the collected dataset run the following command:

python train_skill_modules.py --config_file block/config.yaml --dataset_name fetch_block_40000

To visualise the performance of the trained skill module run the following command:

python utils/test_skill_modules.py --dataset_name fetch_block_40000 --task block --use_skill_prior True

To train the ReSkill agent using the trained skill modules, run the following command:

python train_reskill_agent.py --config_file table_cleanup/config.yaml --datatset_name fetch_block_40000

Logging

All results are logged using Weights and Biases. An account and initial login is required to initialise logging as described on thier website.

Code Structure

reskill
   |-- data 			# collected demonstration data
   |-- reskill			# contains all executable code 
   |   |-- configs 		# all config files for experiments
   |   |   |-- rl  		# config files for rl experiements
   |   |   `-- skill_mdl	# config files for both skill vae and skill prior
   |   |-- data			# dataset specifc code for collection and loading
   |   |-- models		# holds all model classes that implement forward and loss
   |   |-- results		# stores all trained pytorch models for both skill and rl modules
   |   |-- rl			# all code related to rl
   |   |   |-- agents		# implements core algorithms for rl agents
   |   |   |-- envs		# defines the set of environments for data collection and training
   |   |   `-- utils		# utilities for multiprocessing and training distributed rl agents
   |   |-- utils		# general utilities for data management and testing trained modules
   |   |   `-- controllers 	# set of scripted controllers used for data collection
   |   `-- wandb		# logging data from wandb

Citation

  @article{rana2022reskill,
    title={Residual Skill Policies: Learning an Adaptable Skill-based Action Space for Reinforcement Learning for Robotics},
    author={Rana, Krishan and Xu, Ming and Tidd, Brendan and Milford, Michael and S{\"u}nderhauf, Niko},
    journal={Conference on Robot Learning (CoRL) 2022},
    year={2022}
  }

Troubleshooting

ImportError: cannot import name 'PILLOW_VERSION' from 'PIL'

This is due to a Pillow version mismatch between the versions installed by mujoco-py and pytorch. A compatible package that fixes the issue is:

conda install pillow=6.1

---

Creating window glfw
ERROR: GLEW initalization error: Missing GL version

Error when rendering a MuJoCo window for visualisation. This can be fixed by exporting the required package:

export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so