README.md

metadata

task_categories:
  - robotics
tags:
  - lerobot
  - cross-embodiment

MOTIF: Learning Action Motifs for Few-shot Cross-Embodiment Transfer

This repository contains a minimal real-world dataset provided to reproduce the interleaved task setting described in the paper MOTIF: Learning Action Motifs for Few-shot Cross-Embodiment Transfer.

GitHub | Paper

Dataset Description

MOTIF is a framework for few-shot cross-embodiment robotic transfer. It learns reusable action motifs—embodiment-agnostic spatiotemporal patterns—that enable efficient policy generalization across different robot embodiments.

This example dataset includes:

• Embodiments: ARX5 and Piper.
• Tasks: Two distinct tasks across embodiments.
• Format: The dataset adheres to the LeRobot data format and includes a modality.json for detailed modality and annotation definitions (compatible with GR00T N1).

Usage

Download the Dataset

You can download the dataset locally using the huggingface-cli:

huggingface-cli download \
  --repo-type dataset Crossingz/ARX5_Piper_Few_shot_Example \
  --local-dir ./demo_data

Kinematic Trajectory Canonicalization

To enable embodiment-agnostic motif learning, raw end-effector trajectories must be canonicalized into a shared reference frame. You can use the processing script provided in the official repository:

python data/process/trajectory_canonicalization.py \
  --dataset_path ./demo_data \
  --save_path ./demo_data_processed

Citation

If you find this dataset or the MOTIF framework useful, please consider citing:

@article{zhi2025motif,
  title={MOTIF: Learning Action Motifs for Few-shot Cross-Embodiment Transfer},
  author={Zhi, Heng and Tan, Wentao and Zhu, Lei and Li, Fengling and Li, Jingjing and Yang, Guoli and Shen, Heng Tao},
  journal={arXiv preprint arXiv:2602.13764},
  year={2025}
}