News

• 2025/06/12: Manage the project dependency using uv. Isolate the evaluation and training environments
• 2025/06/08: Release the evaluation code and ckpts on LIBERO.
• 2025/06/03: We release the training codebase of Hume.

TODO LIST

• Release Training Code
• Release LIBERO Evaluation
• Switch to uv Manage the Project Dependency
• Release Real-Wold Evaluation
• Release Simpler Evaluation

Try Hume

from hume import HumePolicy
import numpy as np

# load policy
hume = HumePolicy.from_pretrained("/path/to/checkpoints").to("cuda")

# config Test-Time Computing args
hume.init_infer(
    infer_cfg=dict(
        replan_steps=8,
        s2_replan_steps=16,
        s2_candidates_num=5,
        noise_temp_lower_bound=1.0,
        noise_temp_upper_bound=1.0,
        time_temp_lower_bound=0.9,
        time_temp_upper_bound=1.0,
        post_process_action=True,
        device="cuda",
    )
)

# prepare observations
observation = {
    "observation.images.image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C)
    "observation.images.wrist_image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C)
    "observation.state": np.zeros((1, 8)), # (B, state_dim) match the state dimension of the corresponding model
    "task": ["Lift the papper"],
}

# Infer the action
action = hume.infer(observation) # (B, action_dim)

LIBERO

Documents

🛠️ Installation

We use uv manage our project dependency

Install uv in just one line command

curl -LsSf https://astral.sh/uv/install.sh | sh

git clone https://github.com/hume-vla/hume.git
GIT_LFS_SKIP_SMUDGE=1 uv sync

🌟 Train System2

• Download pretrained System2 Weights hume-system2-pt
• Set pretrained_policy in scripts/train_s2.sh to path of the pretrained System2 weights
• Set environment veriable in scripts/env.sh
• Launch training

bash scripts/train_s2.sh

🌟 Train System2 and Value Query Head

• Set pretrained_s2_path in scripts/train_vqh_s1.sh
• Download DinoV2-Small and set --pretrained_dino_path to the DinoV2-Small in scripts/train_vqh_s1.sh
• Launch training

bash scripts/train_vqh_s1.sh

🤗 Model Zoo

Model Name	Link	Note
Hume-System2	hume-system2-pt	pretrained weights exported from pi0

🎄 Datasets

• You can download all lerobot dataset used in our project from here

Dataset Name	Link
LIBERO-Spatial	IPEC-COMMUNITY/libero_spatial_no_noops_lerobot
LIBERO-Goal	IPEC-COMMUNITY/libero_goal_no_noops_lerobot
LIBERO-Object	IPEC-COMMUNITY/libero_object_no_noops_lerobot
LIBERO-10	IPEC-COMMUNITY/libero_10_no_noops_lerobot
Bridge	IPEC-COMMUNITY/bridge_orig_lerobot
Fractal	IPEC-COMMUNITY/fractal20220817_data_lerobot

• If you want to train Hume using your own dataset, check the convert scripts in any4lerobot: 🚀 A collection of utilities and tools for LeRobot.

🤗 FAQs

If you encounter any issues, feel free to open an issue on GitHub or reach out through discussions. We appreciate your feedback and contributions! 🚀

License

This project is released under the MIT license. Parts of this project contain code and models from other sources, which are subject to their respective licenses.

Citation

If you find this project useful in your research, please consider cite:

@article{song2025hume,
  title={Hume: Introducing System-2 Thinking in Visual-Language-Action Model},
  author={Anonimous Authors},
  journal={arXiv preprint arXiv:2505.21432},
  year={2025}
}