| --- |
| language: |
| - en |
| tags: |
| - robotics |
| - embodied-ai |
| - robot-learning |
| - vision-language-action |
| - world-model |
| - lerobot |
| arxiv: "2607.03941" |
| --- |
| |
| # WSA Model Collection |
|
|
| **WSA<sub>1</sub>: a 3D-Centric World-Spatial-Action Model for Generalizable Robot Control** |
|
|
| [Paper](https://arxiv.org/abs/2607.03941) | [Project Page](https://zaleni.github.io/WSA1/) | [Code](https://github.com/zaleni/WSA) | [Model Collection](https://huggingface.co/collections/zaleni/wsa1) |
|
|
| WSA is a robot foundation model built on the 3D-centric **World-Spatial-Action** modeling paradigm. It jointly learns instruction-aligned 2D visual planning, action-conditioned 3D world modeling, and 3D-aware action generation. The released family contains 3B WSA-Base and 6B WSA-Large checkpoints for downstream fine-tuning, RoboTwin2.0, and LIBERO. |
|
|
| ## Released Models |
|
|
| | Model | Size | Checkpoint type | Intended use | |
| | --- | ---: | --- | --- | |
| | [WSA-Base](https://huggingface.co/zaleni/WSA-Base) | 3B | Pretrained | Initialization for downstream fine-tuning | |
| | [WSA-Base-RoboTwin](https://huggingface.co/zaleni/WSA-Base-RoboTwin) | 3B | Fine-tuned | RoboTwin2.0 evaluation and inference | |
| | [WSA-Base-LIBERO](https://huggingface.co/zaleni/WSA-Base-LIBERO) | 3B | Fine-tuned | LIBERO evaluation and inference | |
| | [WSA-Large](https://huggingface.co/zaleni/WSA-Large) | 6B | Pretrained | Initialization for downstream fine-tuning | |
| | [WSA-Large-RoboTwin](https://huggingface.co/zaleni/WSA-Large-RoboTwin) | 6B | Fine-tuned | RoboTwin2.0 evaluation and inference | |
| | [WSA-Large-LIBERO](https://huggingface.co/zaleni/WSA-Large-LIBERO) | 6B | Fine-tuned | LIBERO evaluation and inference | |
|
|
| WSA-Base uses [Qwen3-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct) as its backbone. WSA-Large uses [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) as its backbone. |
|
|
| ## Results |
|
|
| ### RoboTwin2.0 |
|
|
| Average success rate on the randomized (hard) setting over 50 simulated ALOHA manipulation tasks: |
|
|
| | Model | Average Success (Hard) | |
| | --- | ---: | |
| | WSA<sub>1</sub>-B | 92.70% | |
| | **WSA<sub>1</sub>-L** | **93.14%** | |
|
|
| ### LIBERO |
|
|
| Success rates (%) on the four LIBERO task suites: |
|
|
| | Method | LIBERO-Spatial | LIBERO-Object | LIBERO-Goal | LIBERO-10 | Average | |
| | --- | ---: | ---: | ---: | ---: | ---: | |
| | WSA<sub>1</sub>-B | 98.6 | 99.6 | 97.2 | 94.2 | 97.4 | |
| | **WSA<sub>1</sub>-L** | **99.4** | **99.8** | **98.0** | **95.6** | **98.2** | |
|
|
| Refer to the [paper](https://arxiv.org/abs/2607.03941) and [project page](https://zaleni.github.io/WSA1/) for the complete experimental setup, baseline details, and real-world results. |
|
|
| ## Citation |
|
|
| If you use WSA in your research, please cite: |
|
|
| ```bibtex |
| @misc{jiang2026wsa1, |
| title = {WSA$_1$: a 3D-Centric World-Spatial-Action Model for Generalizable Robot Control}, |
| author = {Jiahao Jiang and Jianing Zhang and Zhenhan Yin and Ruidong Chen and Sen Wang and Zhaoshu Yu and Pengpeng Zeng and Xiaofeng Cao and Xuanhan Wang and Jingkuan Song and Heng Tao Shen}, |
| year = {2026}, |
| eprint = {2607.03941}, |
| archivePrefix = {arXiv}, |
| primaryClass = {cs.RO}, |
| url = {https://arxiv.org/abs/2607.03941} |
| } |
| ``` |
|
|
| ## Acknowledgments |
|
|
| WSA builds on the open-source efforts of LeRobot, Qwen3-VL, Wan2.2, Cosmos Tokenizer, Depth Anything 3, RoboTwin, LIBERO, InternVLA-A1, and Fast-WAM. Please also follow the licenses and citation requirements of the corresponding projects. |
|
|
