🚀 联合发布：清华大学 (Tsinghua University) & 智源人工智能研究院 (Beijing Academy of Artificial Intelligence - BAAI)

This repository contains the official implementation of MOVE (MOtion-Based Variability Enhancement), a novel and simple data collection paradigm that significantly enhances spatial generalization and data efficiency for robotic manipulation tasks through dynamic demonstrations.

Imitation learning methods in robotics are often constrained by data scarcity and a lack of spatial generalization. Standard data collection relies on static demonstrations, where the environment's spatial configuration (object, target, camera pose) remains fixed throughout a trajectory. This leads to spatial sparsity—the policy only succeeds around training points.

MOVE addresses this by injecting motion into movable objects (pickup object, target object) and the camera during the expert demonstration process. This augmentation implicitly generates a dense and diverse set of spatial configurations within a single trajectory, leading to:

• Richer Spatial Information: One trajectory encodes information from multiple spatial locations.
• Superior Spatial Generalization: Policies trained with MOVE data generalize much better to unseen spatial configurations.
• Enhanced Data Efficiency: MOVE achieves comparable performance with significantly less data than static methods.

MOVE consistently and significantly outperforms the static data collection paradigm.

In a real-world pick-and-place task with highly randomized configurations, MOVE dramatically reduces the required data scale.

MOVE with 35k timesteps matches the performance of the static method trained with over twice the data (75k).

MOVE implements the following controlled kinematic motions to enrich the training data:

1. Object Translation ($V_{o}$): Linear motion with boundary bounce for the pickup/target object.
2. Object Rotation ($\omega_{i}$): 1-D rotation around the vertical z-axis.
3. Camera Movement ($V_{c}$): Camera moves along a constrained cylindrical path relative to the scene.

In multi-phase tasks like Pick-Place, the motion augmentation is applied strategically:

• Pick Phase: Motion is applied only to the pickup object.
• Placement Phase: Motion is applied only to the target object.

# 克隆仓库
git clone [https://github.com/lucywang720/MOVE.git](https://github.com/lucywang720/MOVE.git)
cd MOVE

bash generate/scripts/gen_demonstration_metaworld.sh 

bash MOVE/diffusion_policy/train.sh

if you find this work helpful, please consider citing our paper:

@misc{wang2025movesimplemotionbaseddata,
      title={MOVE: A Simple Motion-Based Data Collection Paradigm for Spatial Generalization in Robotic Manipulation}, 
      author={Huanqian Wang and Chi Bene Chen and Yang Yue and Danhua Tao and Tong Guo and Shaoxuan Xie and Denghang Huang and Shiji Song and Guocai Yao and Gao Huang},
      year={2025},
      eprint={2512.04813},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.04813}, 
}