Mimic Intent, Not Just Trajectories

An intent-to-execution policy for precise and transferable robotic manipulation.

<a href="https://arxiv.org/abs/2602.08602"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-2602.08602-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white"></a>
<a href="https://huggingface.co/huangrm/MINT-libero"><img alt="Hugging Face Policy" src="https://img.shields.io/badge/HuggingFace-Policy-ffca28?style=for-the-badge&logo=huggingface&logoColor=black"></a>
<a href="https://huggingface.co/huangrm/MINT-tokenizer-libero"><img alt="Hugging Face Tokenizer" src="https://img.shields.io/badge/HuggingFace-Tokenizer-ffca28?style=for-the-badge&logo=huggingface&logoColor=black"></a>







<img alt="Python" src="https://img.shields.io/badge/Python-3.12-3776AB?style=for-the-badge&logo=python&logoColor=white" />
<img alt="Status" src="https://img.shields.io/badge/Status-In%20Progress-f59e0b?style=for-the-badge" />
<img alt="License" src="https://img.shields.io/badge/License-MIT-16a34a?style=for-the-badge" />

SDAT maps each action chunk into multi-scale tokens: coarse tokens capture intent, and fine tokens capture execution details. The S1 token space forms behavior-level clusters.

MINT predicts tokens from intent to execution with next-scale autoregression, then decodes them into actions. Intent-based ensemble improves long-horizon stability.

Overview ✨

We introduce MINT (Mimic Intent, Not just Trajectories), a framework for end-to-end imitation learning in dexterous manipulation. MINT explicitly disentangles behavior intent from execution details by learning a hierarchical, multi-scale token representation of actions. Coarse tokens capture global, low-frequency intent, while finer tokens encode high-frequency execution details. Our policy generates trajectories via next-scale autoregression, performing progressive intent-to-execution reasoning. This structure enables efficient learning, robust adaptation to environmental dynamics, and one-shot skill transfer by reusing the intent token from a demonstration. Experiments on simulation and real robots demonstrate strong performance, high generalization, and effective skill transfer.

Open-Source Roadmap 🗺️

Installation 🛠️

conda create -y -n mint python=3.12
conda activate mint

pip install lerobot==0.4.3
# Install LIBERO dependencies via LeRobot:
pip install "lerobot[libero]==0.4.3"

conda install -y ffmpeg -c conda-forge

cd lerobot_policy_mint
pip install -e .

Note: If you encounter build errors on Linux, you may also need system packages such as cmake, build-essential, python3-dev, pkg-config, and FFmpeg development libraries.

apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev

Model Zoo 🧩

MINT-4B (Lerobot implementation) 🤗

MINT-30M (A lightweight implementation) ⚡

Training Example 🏋️

First, download the required tokenizer:

hf download huangrm/MINT-tokenizer-libero --local-dir <path/to/tokenizer>

Or, train your own tokenizer:

# install tokenizer training dependencies
pip install -r requirements.txt
python -m SDAT.train --config-name train

Start training:

lerobot-train \
    --dataset.repo_id=HuggingFaceVLA/libero \
    --policy.type=mint \
    --output_dir=./outputs/mint_training \
    --job_name=mint_training \
    --policy.repo_id=mint \
    --policy.pretrained_path=huangrm/pi05_base \
    --policy.vqvae_name_or_path=<path/to/tokenizer> \
    --policy.compile_model=false \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --steps=10000 \
    --policy.device=cuda \
    --batch_size=32

Evaluation 📊

lerobot-eval \
    --policy.path=huangrm/MINT-libero \
    --policy.vqvae_name_or_path=<path/to/tokenizer> \
    --env.type=libero \
    --env.task=libero_10,libero_object,libero_spatial,libero_goal \
    --eval.batch_size=1 \
    --eval.n_episodes=2 \
    --policy.n_action_steps=4

Citation 📚

If you find this project useful, please cite:

@article{huang2026mimic,
  title={Mimic Intent, Not Just Trajectories},
  author={Huang, Renming and Zeng, Chendong and Tang, Wenjing and Cai, Jintian and Lu, Cewu and Cai, Panpan},
  journal={arXiv preprint arXiv:2602.08602},
  year={2026}
}

Acknowledgement 🙏

This project is built on top of excellent open-source ecosystems. We sincerely thank the teams behind LeRobot and OpenPI for their impactful contributions.