--- license: apache-2.0 tags: - vla - iclr - iclr-2026 - vision-language-action - spatial-understanding - generalist-robot-policies - calvin-dataset ---

| FALCON | From Spatial to Actions:
Grounding Vision-Language-Action Model in Spatial Foundation Priors (ICLR 2026)

Zhengshen Zhang Hao Li Yalun Dai Zhengbang Zhu Lei Zhou
Chenchen Liu Dong Wang Francis E. H. Tay Sijin Chen
Ziwei Liu Yuxiao Liu^*^† Xinghang Li^* Pan Zhou^*

^*Corresponding Author ^†Project Lead

ByteDance Seed
National University of Singapore Nanyang Technological University
Tsinghua University Singapore Management University

## 🚀 Introduction Existing vision-language-action (VLA) models act in 3D real-world but are typically built on 2D encoders, leaving a spatial reasoning gap that limits generalization and adaptability. In this work, we introduce **FALCON (From Spatial to Action)**, a novel paradigm that injects rich 3D spatial tokens into the action head of a VLA model, enabling robust spatial understanding and SOTA performance across diverse manipulation tasks without disrupting vision-language alignment. See our paper at [here](https://arxiv.org/abs/2510.17439). # Dataset Card for CALVIN Dataset_ABC-D in Point Cloud Format ## Dataset Details This dataset repo contains preprocessed point cloud data in world coordinate system from the static camera and wrist camera under the ABC-D setting (training/validation) of the [Calvin dataset](https://github.com/mees/calvin). **NOTE:** The related camera extrinsic parameters are deprecated. For accurate static camera parameters, please refer to [this repo](https://huggingface.co/datasets/FALCON-VLA/CALVIN-3D_cam-params). ### Dataset Sources - **Repository:** [CALVIN Dataset_ABC-D](https://github.com/mees/calvin) - **Paper:** [CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks](https://arxiv.org/pdf/2112.03227) ## 📦 Usage We provide an efficient way to download the dataset in shards, merge them into a single archive, and then extract the final data. 1. Prepare a folder for downloaded parts ```bash mkdir -p downloaded_parts/ ``` 2. Download the dataset shards You can use either the official `Hugging Face CLI` or the `hfd.sh` script: **Option A:** Hugging Face CLI ```bash huggingface-cli download FALCON-VLA/CALVIN-3D_PCD-ABC_D --repo-type dataset ``` **Option B:** High-speed download with hfd.sh (recommended) ```bash # Install dependencies sudo apt-get install aria2 git-lfs -y # Download the helper script wget https://hf-mirror.com/hfd/hfd.sh chmod +x hfd.sh # Download the dataset shards ./hfd.sh FALCON-VLA/CALVIN-3D_PCD-ABC_D --dataset --tool aria2c -x 8 -j 5 --include "*.tar.gz" ``` 3. Merge the downloaded shards Make sure all shard files are placed under `downloaded_parts/`, then merge them into a single tarball: ```bash cat downloaded_parts/packaged_ABC_D*.tar.gz > packaged_ABC_D.tar.gz ``` 4. Extract the merged archive ```bash tar -xzvf ./packaged_ABC_D.tar.gz -C ./ ``` After extraction, the dataset files will be available in the current directory. For point cloud data loading details, please refer to our released data class [**DiskCalvinDataset3D**](https://github.com/FALCON-VLA/FALCON/blob/92d466736d1dd5cdcbb62a27277f0a4113273334/falcon/data/calvin_dataset.py#L1045). ## Dataset Structure After extraction, the dataset is organized as follows: ```bash packaged_ABC_D/ ├── training/ │ ├── episode_0000001.npz │ ├── episode_0000002.npz │ └── ... └── validation/ ├── episode_0000038.npz ├── episode_0000039.npz └── ... ``` Each .npz file corresponds to one episode and contains the following fields: ```bash static_pcd: point cloud from the static camera view, shape: (200, 200, 3) static_rgb: RGB image from the static camera view, shape: (200, 200, 3) gripper_pcd: point cloud from the wrist camera view, shape: (84, 84, 3) gripper_rgb: RGB image from the wrist camera view, shape: (84, 84, 3) annotation_id: episode annotation identifier, type: int ``` **Note:** `static_cam_ex_mat` and `gripper_cam_ex_mat` are deprecated and are no longer used in the current pipeline. ## 🤗 FAQs If you encounter any issues, feel free to open an issue or reach out through discussions. We appreciate your feedback and contributions! 🚀 ## 🖊️ Citation If you find this project useful in your research, please consider cite: ```BibTeX @article{zhang2025spatial, title={From spatial to actions: Grounding vision-language-action model in spatial foundation priors}, author={Zhang, Zhengshen and Li, Hao and Dai, Yalun and Zhu, Zhengbang and Zhou, Lei and Liu, Chenchen and Wang, Dong and Tay, Francis EH and Chen, Sijin and Liu, Ziwei and others}, journal={arXiv preprint arXiv:2510.17439}, year={2025} } ``` ```BibTeX @article{mees2022calvin, title={Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks}, author={Mees, Oier and Hermann, Lukas and Rosete-Beas, Erick and Burgard, Wolfram}, journal={IEEE Robotics and Automation Letters}, volume={7}, number={3}, pages={7327--7334}, year={2022}, publisher={IEEE} } ``` ## 🪪 License All datasets, as well as our [codebase](https://github.com/FALCON-VLA/FALCON) are released under the Apache-2.0 License. ## ❤️ Acknowledgement FALCON is built with reference to the code of the following projects: [RoboVLMs](https://github.com/Robot-VLAs/RoboVLMs/tree/main?tab=readme-ov-file), [Microsoft Kosmos-2](https://github.com/microsoft/unilm/tree/master/kosmos-2), [VGGT](https://github.com/facebookresearch/vggt), and [ManiUniCon](https://github.com/Universal-Control/ManiUniCon). Thanks for their awesome work!

| FALCON | From Spatial to Actions:Grounding Vision-Language-Action Model in Spatial Foundation Priors (ICLR 2026)

| FALCON | From Spatial to Actions:
Grounding Vision-Language-Action Model in Spatial Foundation Priors (ICLR 2026)