starvla/starvla

StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing

An open-source research platform for integrating and exploring cutting-edge technologies for generalist robots.

📢 Citation Update: Our technical report is now on arXiv (2604.05014). We kindly invite you to use the updated BibTeX for any ongoing or future citations. If you have already cited StarVLA in a previous version of your work, we would greatly appreciate it if you could update the citation entry in your camera-ready or future revisions. Thank you for your understanding and support! 🙏

---

In StarVLA (also a pun on “start VLA” ), each functional component (model, data, trainer, config, evaluation, etc.) follows a top-down, intuitive separation and high-cohesion, low-coupling principle, enabling plug-and-play design, rapid prototyping, and independent debugging.

Our Team’s Work

News

⚠️ Branch notice: The starVLA_dev branch is where we actively merge new features and may be temporarily unstable. For verified results, use the stable starVLA branch. Thanks to StarVLA’s low-coupling design, switching between branches is painless. We encourage trying starVLA_dev and welcome PRs if you spot any issues!

💡 Tip: Files under any **/bar/ directory are git-ignored, so you can place your custom scripts there (e.g., examples/simBenchmarks/LIBERO/train_files/bar/my_train.sh) without polluting the repo.

[2026/09/05] 🏠 We have updated the RoboCasa365 (PandaOmron) integration with a consolidated data, training, and evaluation guide and a 50-task local research snapshot covering StarVLA-OFT, StarVLA-PI, and StarVLA-GR00T. The best selected local result is StarVLA-PI at 40.64% (1,016/2,500). These target-task/target-kitchen results are provided for research reference and are not official RoboCasa leaderboard submissions.

[2026/08/27] 🚀 We released VLAct and its paper, Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models. VLAct introduces representation-centric continued pre-training for reusable VLA backbones; the code and models are now public. If you already use a Qwen3-VL-4B StarVLA recipe, set framework.qwenvl.base_vlm to StarVLA/VLAct_Qwen3_Pretrain to try the VLAct-pretrained backbone directly.

[2026/08/09] 🤖 StarVLA now supports RoboDojo through XPolicyLab. The example provides training recipes for the official RoboDojo dataset and a path-based entry point for evaluating the released checkpoints with XPolicyLab.

[2026/06/22] 🔥 We are creating a StarVLA Contributors Group to make contributor communication easier and discuss contributor list before each release. Welcome everyone to help maintain StarVLA’s open-source infrastructure together.

[2026/05/30] 🔥 StarVLA now supports VLA training with Qwen-series backbones on Ascend NPU. See the related discussion for details and feedback.

[2026/04/09] 🚀 unified multi-benchmark co-training example (combining LIBERO, SimplerEnv, RoboTwin, VLA-Arena, etc.) is coming soon. Stay tuned!

[2026/05/01] 🔥 We now provide a step-by-step guide for integrating your own robot / dataset into StarVLA.

[2026/05/01] 🔥 We are building agent skills to make StarVLA a powerful substrate for AI coding agents — we have verified that GitHub Copilot (Claude Opus 4.7) can autonomously integrate examples/Robocasa_365 and examples/RoboChallenge_table30v2 from scratch. Going forward, StarVLA will be continuously optimised to be equally easy to use for humans and code agents.

[2026/05/01] 🙏 Special thanks to the AllenAI team for providing optimised Docker environments and evaluation-acceleration support via vla-evaluation-harness! If you need to speed up large-scale VLA benchmark evaluation, we recommend checking it out.

[2026/04/18] 🔥 StarVLA now supports DOMINO, a dynamic manipulation benchmark for moving objects and time-varying scenes. Original DOMINO repository is here.

[2026/04/09] 🎯 Thanks to the RLinf team, StarVLA now supports RL post-training! Check out the StarVLA × RLinf tutorial to get started.

[2026/04/09] 🔥 WM4A (World Model for Action) is now integrated! Use pretrained video-generation DiT models (Cosmos-Predict2, Wan2.2) as backbones for action prediction. See docs/WM4A.md for architecture details and training instructions. Pretrained checkpoints are available at the StarVLA/world-model-to-vla HuggingFace collection.

[2026/03/29] 🔥 Thanks to the ABot-M0 team for providing the pre-trained weights. For Qwen3-VL 4B, you can reload the qwen_vl_interface module in various frameworks!

[2026/03/19] 🔥 StarVLA now provides a complete real-robot development case with Franka robot examples!

[2026/03/03] 🔥 We now support Qwen3.5 as a backbone for VLA — the fastest integration in the community ⚡ With more model size options: 0.8B, 2B, 4B, and 9B! Build your VLA flexibly on top of native multimodal models!

[2026/01/29] 🔥 StarVLA Training Efficiency Report & Training Curves released! Training configs and efficiency benchmarks for community reference.

[2026/01/29] Calvin benchmark experiments were conducted by the UNT team. For inquiries, please contact Zhijie Song ([email protected]) or Feng Yan ([email protected]).

[2025/12/25] We’ve simultaneously established pipelines for Behavior-1K, RoboTwin 2.0, and CALVIN. We’d love to collaborate and share baseline results for more benchmarks with the community!

Overview and Key Features

Overview of the StarVLA framework. StarVLA organises VLA research as a composable stack: a shared training infrastructure, pluggable foundation-model backbones (VLM / world model), interchangeable action heads (FAST, OFT, flow-matching π, GR00T-style dual-system), and benchmark-agnostic deployment hooks. Each axis is decoupled, so a new framework variant typically reduces to swapping the backbone or the action head while reusing the rest.

---

🎒 Quick Start

📖 New to StarVLA? Check out our step-by-step Quick Start Guide — a complete walkthrough from installation to training to evaluation using the LIBERO benchmark.

---

Benchmark Results

We have more results for RoboCasa, RoboTwin 2.0, Behavior-1k, Calvin. See our 🍀 Overleaf, which continuously presents our real-time experimental results.

---

Model Zoo

See the full list of released models and checkpoints in docs/model_zoo.md.

---

Start Building Your VLA Like Lego!

👇 StarVLA achieves “Lego-like” development via the following designs:

---

FAQ

See docs/faq.md for common questions on configuration, freezing, learning rates, checkpointing, smaller VLMs, and more.

Contributing

Community contributors are the driving force behind StarVLA’s growing ecosystem. We deeply appreciate every PR, bug fix, and piece of feedback from the open-source community — your efforts keep StarVLA evolving rapidly. A full, continuously updated contributor list is maintained at starvla.github.io/contributors.

Thanks to all the people who have contributed to StarVLA:

See docs/CONTRIBUTING.md for guidelines on reporting bugs, proposing features, and submitting PRs.

Projects Based on StarVLA

NeuroVLA: A Brain-like Embodied Intelligence for Fluid and Fast Reflexive Robotics Control

PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence

TwinBrainVLA: TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers

LangForce (ICML 2026): LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries

LaWAM: LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

Examples:

accelerate launch \
  --config_file starVLA/config/deepseeds/deepspeed_zero2.yaml  \
  --num_processes 8 \
  starVLA/training/train_internvla.py \
  --config_yaml examples/simBenchmarks/SimplerEnv/train_files/starvla_cotrain_oxe.yaml \
  --framework.qwenvl.base_vlm Qwen/Qwen2.5-VL-7B-Instruct \ # override framework choice
  --framework.qwenvl.base_vlm Qwen/Qwen2.5-VL-7B-Instruct \ # override framework choice
  --framework.action_model.new_module ${module_name} \ # plug-in a new module to action model

⚠️: framework.action_model.new_module only adds to the global config; its behavior is on your framework.

✍️ Citation & Copyright

StarVLA is released under the MIT License, which permits commercial use, modification, distribution, and private use. Rebases are allowed for forks and feature branches; when rebasing from upstream StarVLA, use descriptive commit messages (e.g., “chore: rebase from StarVLA”) and keep at least the two latest upstream commits as separate. See License for details.

@article{yang2026beyond,
  title={Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models},
  author={Yang, Senqiao and Wang, Chengyao and Chen, Yuxin and Wang, Zixuan and Tang, Longxiang and Gui, Haokun and Ye, Jinhui and Lu, Changsheng and Wu, Xiaoyang and Zhu, Mingkang and others},
  journal={arXiv preprint arXiv:2608.27550},
  year={2026}
}

@inproceedings{ye2026starvla,
  title={StarVLA-$$\backslash$alpha $: Reducing Complexity in Vision-Language-Action Systems},
  author={Ye, Jinhui and Gao, Ning and Yang, Senqiao and Zheng, Jinliang and Wang, Zixuan and Chen, Yuxin and Chen, Pengguang and Chen, Yilun and Liu, Shu and Jia, Jiaya},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

@article{community2026starvla,
  title={StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing},
  author={Community, StarVLA},
  journal={arXiv preprint arXiv:2604.05014},
  year={2026},
  eprint={2604.05014},
  archivePrefix={arXiv},
  primaryClass={cs.RO}
}

Acknowledgements

This project draws inspiration and references from several notable open-source initiatives, including:

• LeRobot
• GR00T
• DeepSpeed
• Qwen-VL
• InternVL
• ABot-Manipulation

The codebase was originally forked from InternVLA-M1.

Star History

Here’s how our community has grown over time: