Neil Janwani, Ellen Novoseller, Vernon Lawhern, Maegan Tucker,

https://arxiv.org/abs/2603.09237v1

MO-Playground is a collection of multi-objective environments built in JAX for GPU-Accelerated multi-objective RL.

Note that due to double-blind requirements, moplayground's documentation page and pip-installable package is not yet available.

The code was tested with:

• Ubuntu 22.04
• Python 3.12.12
• CUDA 13.0 (if you want to train policies. Evaluation can happen without a GPU).

Using the provided ymls, create a new conda environment. If you want to enable GPU based training, run

conda env create -f environment.yml
conda activate moplayground

if you just want to evaluate policies and explore the code (i.e. running on a Mac)

conda env create -f mac_environment.yml
conda activate moplayground

Finally, go to the project root and run

pip3 install -e .

to install the moplayground package.

Create an account at Weights and Biases. Note that the process should be free and you'll be required to paste your API key to get things to work. Educational accounts receive some amount of free storage by the way, which can be useful if you're a student!

Finally, pick an environment from the below list

• cheetah
• hopper
• walker
• ant
• humanoid
• bruce

and download your desired policy.

python3 -m scripts.download_model --env cheetah  

note that you can supply a desired save directory via --save_dir. The default directory is simply results/wandb-downloads. Finally, you can run the policy via

python3 -m scripts.rollout_policy config_path

where config_path is the config.yaml file where your model was saved. It will be at save_dir/env_name/config.yaml, where save_dir and env_name are defined above.

To train a pre-existing environment, check out the configuration files in config/. These files specify everything from model architecture and MORLAX parameters to reward and environment constants. Choose the config file you want, edit the parameters to your liking, and run

python3 -m scripts.train config_path

where config_path is the path to the config of your choice. If you downloaded a policy in the past, you can also use those configs to run an identical training run on your system.

To create a custom environment, check out how the cheeah environment works at src/moplayground/envs/dmcontrol/cheetah.py. You should need to make your child class a member of the MultiObjectiveBase class. You will also need to make a config.yaml file for your environment to specify the training parameters.

Note that support for custom dynamics (i.e. non-mujoco) is coming soon.

Environment	Reward 1	Reward 2

MO-Playground is demonstrated for the BRUCE humanoid robot, developed by Westwood Robotics.

The application features seven possible reward functions. Note that we combine base_xyz_tracking and base_quat_tracking to explore a 6-dimensional objective space.

Policy	Result
Balanced Reward
Max Imitation
Max Arm Swinging
Max Smoothness

@article{janwani2026mo,
  title={MO-Playground: Massively Parallelized Multi-Objective Reinforcement Learning for Robotics},
  author={Janwani, Neil and Novoseller, Ellen and Lawhern, Vernon J and Tucker, Maegan},
  journal={arXiv preprint arXiv:2603.09237},
  year={2026}
}