Robust Reinforcement Learning Suite (rrls)

Goal

The goal of rrls is to standardize robust reinforcement learning benchmarks, ensuring that experiments are reproducible and comparable. rrls is designed to follow the gymnasium API.

---

📦 Installation

From source:

# We have to install the latest version of gymnasium
git clone https://github.com/Farama-Foundation/Gymnasium.git

cd Gymnasium

pip install .

pip install git+https://github.com/SuReLI/RRLS.git

Available when Gymasium 1.0 is released

Via pip:

pip install rrls

Prerequisites:

• Ensure you have MuJoCo installed on your machine. The environments provided by rrls require the MuJoCo physics engine from Deepmind. For detailed installation instructions, please refer to the MuJoCo website and the MuJoCo Github repository.
• We have tested and support Python versions 3.9, 3.10, and 3.11 on both Linux and macOS.

🤖 Environments

The package offers the following environments:

Environment Name	id
Ant	robust-ant-v0
HalfCheetah	robust-halfcheetah-v0
Hopper	robust-hopper-v0
HumanoidStandup	robust-humanoidstandup-v0
InvertedPendulum	robust-invertedpendulum-v0
Walker2d	robust-walker2d-v0

And lot more ... if you want to get a full list of the environments, you can use the following code:

import gymnasium as gym
for env in gym.envs.registry:
    if "rrls/robust" in env:
        print(env)

Example of usage:

import gymnasium as gym
import rrls

env = gym.make("rrls/robust-ant-v0")
params = env.get_params()  # Parameters will be None if reset method hasn't been called.

# Set all parameters to 1
params = {k: 1 for k in params.keys()}

# Modify environment parameters during reset using the options argument
obs, info = env.reset(options=params)

# Retrieve environment parameters from the info dictionary
print(info["torso_mass"])  # Outputs: all keys are equals to one 1

terminated, truncated = False, False

while not (terminated or truncated):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    print(info["torso_mass"])

    # To change environment parameters during an episode, use the set_params method.
    # env.set_params(params)

🌯 Wrappers

The package provides the following wrappers:

• Domain randomization: rrls.wrappers.DomainRandomization
• Probabilistic action robustness: rrls.wrappers.ProbabilisticActionRobust
• Adversarial dynamics: rrls.wrappers.DynamicAdversarial

👝 Uncertainty sets

For each environment, we offer a set of uncertainty sets for use. For instance:

from rrls.envs.ant import AntParamsBound

This Enum includes three variants: 1D, 2D, and 3D uncertainty sets, as referenced from the M2TD3 paper. For instance, the 2D uncertainty set for the Ant environment is defined as follows:

from rrls.envs.ant import AntParamsBound

params_bound_2d = AntParamsBound.TWO_DIM.value
# {'torsomass': [0.1, 3.0], 'frontleftlegmass': [0.01, 3.0]}
#                ^^^  ^^^
#                min  max

Also you can get the uncertainty set provided by the RARL paper

from rrls.envs.ant import AntParamsBound
rarl_params_bound = AntParamsBound.RARL.value
# {
#         "torsoforce_x": [-3.0, 3.0],
#         "torsoforce_y": [-3.0, 3.0],
#         "frontleftlegforce_x": [-3.0, 3.0],
#         "frontleftlegforce_y": [-3.0, 3.0],
#         "frontrightlegforce_x": [-3.0, 3.0],
#         "frontrightlegforce_y": [-3.0, 3.0],
#     }

🤓 Evaluate

If you want benchmark worst-case performance using our extensive suite. For every uncertainty set, we provide a corresponding set of evaluation environments. These environments are created by equally partitioning (into 10 segments) each dimension of the uncertainty set.

from rrls.evaluate import EVALUATION_ROBUST_ANT_3D # Set consisting of 10^3 environments

If you wish to construct your own custom set of environments, you can utilize the code below:

from rrls.evaluate import generate_evaluation_set
from rrls.envs.ant import AntParamsBound, RobustAnt

eval_env_set = generate_evaluation_set(
    modified_env=RobustAnt,
    param_bounds=AntParamsBound.THREE_DIM.value,
    nb_mesh_dim=3,
)

📖 Project Maintainers

• Adil Zouitine - IRT Saint-Exupery, ISAE Supaero, & Sureli Team
• David Bertoin - IRT Saint-Exupery, INSA Toulouse, ISAE Supaero, & Sureli Team
• Emmanuel Rachelson - ISAE Supaero & Sureli Team
• Pierre Clavier - Ecole polytechnique and Inria Paris

🙏 Acknowledgments

This project is part of the ENVIA initiative, aiming to develop next-gen reinforcement learning controllers for airship transportation of heavy loads. We are grateful to our supporters:

• Sureli Team
• Isae Supaero
• IRT Saint Exupéry
• Thales