README.md

metadata

license: mit
task_categories:
  - robotics
tags:
  - pick-and-place
  - manipulation
  - reinforcement-learning

This is the official repository for the training dataset of the paper: Efficient Alignment of Unconditioned Action Prior for Language-conditioned Pick and Place in Clutter.

We study the task of language-conditioned pick and place in clutter, where a robot should grasp a target object in open clutter and move it to a specified place. This dataset supports the A$^2$ action prior alignment method, which integrates foundation priors from vision, language, and action to enable effective policies.

Project Page: https://xukechun.github.io/papers/A2 Code: https://github.com/xukechun/Action-Prior-Alignment

Please download the file and unzip it in the data folder.

Sample Usage

The following snippets from the GitHub repository demonstrate how to use this dataset for data collection, training, and evaluation.

Data Collection

• For pick data

bash scripts/data_collection/collect_data_grasp.sh

• For place data

bash scripts/data_collection/collect_data_place.sh

Training

• Unified training for pick and place

bash scripts/train/train_clutter_gp_unified.sh

• Adaptation for place

bash scripts/train/train_clutter_gp_adaptive.sh

Evaluation

To test the pre-trained model, simply change the location of --model_path:

• Pick

bash scripts/test/test_grasp.sh

• Place

bash scripts/test/test_place.sh

• Pick and place

bash scripts/test/test_pickplace.sh

Citation

If you find this work useful, please consider citing:

@article{xu2025efficient,
  title={Efficient Alignment of Unconditioned Action Prior for Language-conditioned Pick and Place in Clutter},
  author={Xu, Kechun and Xia, Xunlong and Wang, Kaixuan and Yang, Yifei and Mao, Yunxuan and Deng, Bing and Xiong, Rong and Wang, Yue},
  journal={arXiv preprint arXiv:2503.09423},
  year={2025}
}