rsl_usr

This is the codebase for CoRL 2023 paper: Robust reinforcement learning in continuous control tasks with uncertainty set regularization.

If you use any source codes included in this toolkit in your work, please cite the following paper. The bibtex is listed below:

    @inproceedings{
    zhang2023robust,
    title={Robust Reinforcement Learning in Continuous Control Tasks with Uncertainty Set Regularization},
    author={Yuan Zhang and Jianhong Wang and Joschka Boedecker},
    booktitle={7th Annual Conference on Robot Learning},
    year={2023},
    url={https://openreview.net/forum?id=keAPCON4jHC}
    }

The code implementation largely depends on a Github Repository Soft Actor-Critic implemtation in Pytorch.

1. Installation

Tested on Ubuntu:20.04 and Python:3.7

Mujoco 2.1.0

Download Mujoco:2.1.0 and extract it into ~/.mujoco/mujoco210

glew

For conda users, please refer https://anaconda.org/conda-forge/glew

conda install -c conda-forge glew

realwolrdrl_suite

cd envs/realworldrl_suite
pip install .

other requirements

pip install -r requirements.txt

2. Instructions

2.1 Training

To train an SAC agent with Uncertainty Set Regularizer introduced in the paper, run scripts/train.sh:

# mujoco 210
export MUJOCO_GL=egl
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco210/bin

# set up cuda 
export CUDA_DEVICE_ORDER=PCI_BUS_ID

# select exp setting and regularizer setting
override=quadruped_walk
robust_method=l2_adv_param
robust_coef=1e-4
exp_name=adv

# run training on 5 random seeds
cuda_id=0
for seed in 12345 23451 34512 45123 51234; do
    # set up cuda
    export CUDA_VISIBLE_DEVICES=${cuda_id}
    cuda_id=$(($cuda_id+1))
    # run training
    python train.py \
        overrides=${overrides} \
        seed=${seed} \
        agent.params.robust_method=${robust_method} \
        agent.params.robust_coef=${robust_coef} \
        experiment=exp_name &
done

• agent.params.robust_method is the method to perform additional regularizers. Support methods:

  • none: no regularizer on value function, original SAC algorithms
  • l1_reg: $l_1$ regularizer on value function 's parameters
  • l2_reg: $l_2$ regularizer on value function 's parameters
  • l1_param: : $l_1$ uncertainty set regularizer on value function
  • l2_param: $l_2$ uncertainty set regularizer on value function
  • l2_adv_param: adversarial uncertainty set regularizer on value function

• agent.params.robust_coef is the robust magnitude hyperparameter for robust RL

• experiment is the special name for this experiment

All experiments are runned on 5 random seeds and saved in ouputs folder. One can run tensorboard to monitor training by running tensorboard --logdir outputs

2.2 Testing

To test a trained agent under perturbatoins, run bash scripts/test_robust.sh

• perturb_param_list is the allowed perturbation parameters in the certain environment
• perturb_min(max)_list restricts the range of the changed vlaue

3. Code Structure

├── __init__.py
├── agents
│   ├── __init__.py
│   ├── actor.py
│   ├── critic.py
│   └── sac.py: implementation of robust SAC agent, including all robust update methods
├── configs
│   ├── agent
│   ├── overrides: parameters of different tasks
│   └── train.yaml
├── envs
│   ├── realworldrl_suite
│   ├── rwrl_env.py: main entry of RWRL benchmark
│   └── toy_env.py: main entry of toy environment
├── logger.py
├── outputs: folder to save experimental results
├── replay_buffer.py
├── requirements.txt
├── scripts: folder to run training/testing
├── test.py: script to test agents under environmental perturbations
├── train.py: script to train agents
├── utils.py
└── video.py: script to render videos of simulation process