High-Quality Open-Source Codebases for Embodied Intelligence
StarRobotics is an open-source organization dedicated to advancing embodied intelligence through modular, production-quality codebases. We believe the future of robotics lies in open collaboration — every component, from perception to action, should be transparent, reproducible, and easy to build upon.
"Making embodied intelligence development as easy as building with Lego blocks."
| Repository | Focus | Status | Description |
|---|---|---|---|
| StarVLA | Vision-Language-Action | ✅ Active | Lego-like codebase for VLA model developing. Supports Qwen-FAST, Qwen-OFT, Qwen-PI, Qwen-GR00T and more. |
| StarWA | World-Action Model | 🚧 In Progress | World model–driven action prediction. Combines visual world simulation with embodied action planning. |
| StarBot | Generic Embodied Robot | 🚧 In Progress | Unified robot control stack for diverse embodiments — manipulation, locomotion, and beyond. |
StarRobotics/
├── StarVLA # VLA: vision + language → action
├── StarWA # WA: world model → action planning
└── StarBot # Bot: generic embodied robot platform
A Lego-like Codebase for Vision-Language-Action Model Developing
StarVLA provides a modular framework for developing VLA models with plug-and-play components. Each functional module (model, data, trainer, config, evaluation) follows high cohesion & low coupling principles.
Highlights:
- 4 VLA frameworks: Qwen-FAST, Qwen-OFT, Qwen-PI, Qwen-GR00T
- Multiple VLM backbones: Qwen2.5-VL, Qwen3-VL, Florence-2
- Benchmark support: SimplerEnv, LIBERO, RoboCasa, and more
- VLM co-training for multimodal multitask learning
- DeepSpeed ZeRO-2/3 & multi-node training
- Stand up a new VLA framework in < 3 hours
📦 GitHub · 🤗 Models & Data
Learning to Act through World Simulation
StarWA focuses on world model–driven action prediction: learning a predictive model of the environment and leveraging it for embodied decision-making. By combining visual world simulation with action planning, StarWA enables robots to reason about consequences before acting.
Planned Features:
- Video prediction / visual world models
- Latent dynamics learning
- Action planning via model-based rollouts
- Integration with StarVLA for hybrid VLA + WA pipelines
- Sim-to-real transfer utilities
📦 Coming Soon
One Stack, Every Robot
StarBot aims to be a unified, robot-agnostic control stack that supports diverse embodiments — from tabletop manipulators to humanoids and mobile platforms.
Planned Features:
- Generic robot interface abstraction
- Multi-embodiment configuration system
- Real-world deployment utilities (policy server, safety layers)
- Cross-embodiment transfer tools
- Hardware drivers & calibration for common platforms
📦 Coming Soon
- StarVLA v1: core VLA framework release
- 4 VLA architectures (FAST, OFT, PI, GR00T)
- SimplerEnv & LIBERO & Robocasa-GR1 & Robotwin 2.0 evaluation
- VLM co-training support
- StarVLA v2: expanded benchmarks & new architectures
- StarWA v0: initial world-action model codebase
- StarBot v0: generic robot platform prototype
- Cross-repo integration (VLA ↔ WA pipelines)
- Real-world deployment toolkit
- Foundation model for general-purpose embodied agents
- Community benchmark & leaderboard
- Sim-to-real bridge across all repos
All StarRobotics codebases share a common set of principles:
| Principle | Description |
|---|---|
| Modular (Lego-like) | Every component is self-contained and independently testable. |
| Explicit Boundaries | Clear interfaces between data, model, and training — no hidden coupling. |
| Flat Configuration | Single config entry point; override anything from CLI. |
| Reproducible | Pin versions, publish checkpoints, ship evaluation scripts. |
| Community-first | MIT-licensed, open roadmap, weekly office hours. |
We welcome contributions across all repos:
- Report bugs — Open an Issue with a minimal repro.
- Propose features — Start a Discussion or fill out the Cooperation Form.
- Submit PRs — Align scope via an Issue first, then
make checkbefore submitting. - Join office hours — Every Friday afternoon for live discussion.
| Channel | Link |
|---|---|
| GitHub Org | github.com/StarRobotics |
| HuggingFace | huggingface.co/StarVLA |
| WeChat Group | Scan the QR code in the StarVLA repo |
| Cooperation Form | Google Form |
@misc{starrobotics2025,
title = {StarRobotics: High-Quality Open-Source Codebases for Embodied Intelligence},
author = {StarRobotics Community},
url = {https://github.com/StarRobotics},
year = {2025}
}
@misc{starvla2025,
title = {StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing},
author = {StarVLA Community},
url = {https://github.com/StarRobotics/StarVLA},
year = {2025}
}All StarRobotics projects are released under the MIT License — free for commercial use, modification, and distribution.
⭐ Star us on GitHub to stay updated!
Building the open-source infrastructure for embodied intelligence, one repo at a time.