EZ-M is a multi-task extension of EfficientZero series focused on humanoid locomotion tasks. We also reserved support for Atari (single-task) and DMControl single-task training. We recommend multi-task training in the future RL due to superior overall sample efficiency. More details could refer to:

Project page, Paper link

The training pipeline is built around Hydra configuration, Ray-based workers, PyTorch models, and a custom C++/Cython Gumbel search backend. The main entry points are:

• ez/train.py for training
• ez/eval.py for evaluation

• MuZero-style training and evaluation workflow
• Support for both discrete and continuous action spaces
• Environment presets under ez/config/exp/
• Distributed execution
• Optional experiment tracking with Weights & Biases
• Support multi-task training with higher overall sample efficiency

• Python 3.8 or later
• CUDA-enabled GPU environment recommended for training
• Dependencies from requirements.txt or requirements_py310.txt

Install dependencies with:

pip install -r requirements.txt

If you are using Python 3.10, you can use:

pip install -r requirements_py310.txt

Before training or evaluation, compile the C++/Cython MCTS module:

cd ez/mcts/ctree_v2
bash make.sh
cd -

Reported scores could be found in results/ez-m-results.json

Run training with one of the experiment configs in ez/config/exp/:

python ez/train.py exp_config=ez/config/exp/dmc_state.yaml

You can also use the provided shell script:

bash scripts/train.sh

Example experiment configs:

• ez/config/exp/atari.yaml
• ez/config/exp/dmc_image.yaml
• ez/config/exp/dmc_state.yaml
• ez/config/exp/maniskill_state.yaml
• ez/config/exp/humanoid_bench_state.yaml

Run evaluation with:

python ez/eval.py exp_config=ez/config/exp/dmc_image.yaml

Or use the script:

bash scripts/eval.sh

ez/
  agents/      Agent definitions and model implementations
  config/      Global and experiment-specific Hydra configs
  data/        Replay buffer, trajectory, and data processing utilities
  envs/        Environment wrappers and integrations
  mcts/        Python and C++/Cython MCTS implementations
  utils/       Training utilities and helper functions
  worker/      Distributed worker logic for training and evaluation
scripts/       Example launch scripts

• Some training scripts include environment-specific settings such as MUJOCO_GL, CUDA_VISIBLE_DEVICES, and wandb login. Adjust them before running on your machine.
• The repository contains multiple experiment presets; choose the one that matches your target environment and observation type.

@article{liu2026scaling,
  title={Scaling Tasks, Not Samples: Mastering Humanoid Control through Multi-Task Model-Based Reinforcement Learning},
  author={Liu, Shaohuai and Ye, Weirui and Du, Yilun and Xie, Le},
  journal={arXiv preprint arXiv:2603.01452},
  year={2026}
}