Unify lego-like framework modules, training recipes, benchmark evaluation, and deployment interfaces so teams can iterate and ship faster. 统一 Lego 式框架模块、训练流程、基准评测与部署接口，让团队更快迭代并交付。
From setup to benchmarking and deployment, the workflow is organized for fast, reliable team iteration. 从环境搭建到基准评测再到部署落地，整体流程围绕团队快速且可靠的迭代而设计。
StarVLA-FAST, StarVLA-OFT, StarVLA-PI, and StarVLA-GR00T share the same trainer, dataloader, and deployment stack for plug-and-play experimentation. StarVLA-FAST、StarVLA-OFT、StarVLA-PI、StarVLA-GR00T 共用训练、数据和部署管线，实现即插即用实验迭代。
Run SimplerEnv, LIBERO, RoboCasa, RoboTwin, and BEHAVIOR with benchmark-specific scripts and clear environment boundaries. 通过基准专用脚本和清晰环境边界，复现 SimplerEnv、LIBERO、RoboCasa、RoboTwin、BEHAVIOR。
Use a unified WebSocket policy interface to bridge simulation and real-robot control without rewriting model serving logic. 使用统一 WebSocket 策略接口连接仿真与真机，无需重写模型服务逻辑。
Choose a goal and copy the command templates. Evaluation workflows are split into two terminals to avoid dependency conflicts. 选择目标后复制命令模板。评测流程默认分为两个终端，以避免依赖冲突。
Numbers below are reported results from the official StarVLA docs and repository, with source links for verification. 以下数字来自 StarVLA 官方文档与仓库公开结果，并附带来源链接供核对。
Percent values reported in the benchmark summary chart. 以下为基准对比图中给出的百分比结果。
The Generalist model is jointly trained across all benchmarks and is included as a reference for the upper-bound potential of our framework. For direct comparisons under similar computational budgets, we focus primarily on Specialist results. Generalist 模型基于全部基准进行联合训练，仅作为展示本框架性能上限潜力的参考。在同等计算预算下的公平对比中，我们主要关注 Specialist 的结果。
| Benchmark 基准 | starVLA Specialist starVLA 专家模型 | starVLA Generalist starVLA 通用模型 | π₀.₅ π₀.₅ |
|---|---|---|---|
| LIBERO | 98.8% | 97.8% | 96.9% |
| LIBERO-Plus | 77.8% | 79.7% | – |
| SimplerEnv WidowX | 64.6% | 70.2% | 46.9% |
| SimplerEnv Google VA | 70.2% | 73.8% | 68.4% |
| SimplerEnv Google VM | 76.0% | 79.3% | 72.7% |
| RoboTwin Clean* | 88.2% | 88.7% | 82.7% |
| RoboTwin Random* | 88.3% | 87.8% | 76.8% |
| RoboCasa-GR1 | 53.8% | 57.3% | 37.0% |
| RoboChallenge | 30.0% | – | 17.7% |
Use published backbones and checkpoints to reproduce or extend StarVLA experiments quickly. 使用已发布骨干和检查点，快速复现或扩展 StarVLA 实验。
Extended vocabulary with fast tokens for discrete action modeling. 扩展 fast token 词表，用于离散动作建模。
Qwen3-VL action-ready backbone for agile downstream experimentation. 面向下游敏捷实验迭代的 Qwen3-VL 动作化骨干。
Bridge + Fractal checkpoint for SimplerEnv-style evaluation. 用于 SimplerEnv 评测的 Bridge + Fractal 检查点。
Get help, follow live progress, and collaborate with researchers and engineers in the StarVLA ecosystem. 获取支持、跟进实时进展，并与 StarVLA 生态中的研究者和工程师协作。
Official setup, benchmarks, and training guides. 官方安装、评测与训练指南。
Source code, issues, pull requests, and updates. 源码、issue、PR 与最新更新。
Published models, checkpoints, and datasets. 已发布模型、检查点与数据集。
Join via issue thread link. 通过 issue 链接加入中文交流群。
Read the StarVLA technical report on arXiv (2604.05014). 在 arXiv 阅读 StarVLA 技术报告 (2604.05014)。
Continuously updated experiment report and analysis. 持续更新的实验报告与分析。
Top-down modular principles for framework extension. 用于框架扩展的自顶向下模块化原则。
Unified policy client-server interface for deployment. 用于部署的一体化策略客户端-服务端接口。
