- 代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_substantive_official_repository_fixed_commit; 最近核验：2026-08-11代码：training and evaluation released
权重：released separately
数据：released separately without license
许可证：NOASSERTION；MIT-derived text plus explicit preservation obligations for two upstream StarVLA commits.
依赖：PyTorch、Transformers、V-JEPA 2.1、StarVLA lineage、LIBERO/RoboTwin
硬件/传感器：4×A100 for LIBERO training、8×B200 for RoboTwin training、bimanual robot for physical deployment
实机证据：Three bimanual tasks, 180 W2-VLA trials
独立复现：None found.
限制/反面证据：Nonstandard additional license obligations；High training hardware；No independent reproduction
当前建议：当前重点·许可证需逐条审查；Substantive code and weights enable inspection, while legal and compute closure remain nontrivial.
重点标签：腕部世界模型主流技术2.0、VLA细粒度操作、可能成为新主流、近期热门
核验来源：来源 1
- 数据 · verified_hold
审核码：verified_hold; 官方关系：verified_downloadable_license_missing; 最近核验：2026-08-11代码：released
权重：released
数据：training data and W2-CoT both downloadable
许可证：not_disclosed；Neither W2-VLA-Training-Data nor W2-VLA-CoT declares a dataset license.
依赖：Hugging Face Hub、W2-VLA preprocessing
硬件/传感器：~27.13GB training data plus ~27MB CoT labels
实机证据：Includes paper training lineage, exact real/sim mixture needs per-file audit
独立复现：None found.
限制/反面证据：Dataset licenses absent；17,323 files make manual provenance audit nontrivial
当前建议：许可证阻断；Downloadability is verified, but reuse permission is not.
重点标签：数据已上传、许可证缺失、重点更新跟踪
核验来源：来源 1 · 来源 2
- 权重 · 已核验官方资源
审核码：verified_official; 官方关系：verified_public_nongated_fixed_revision; 最近核验：2026-08-11代码：released separately
权重：4.97B model package released
数据：released separately without license
许可证：Apache-2.0；Model-card declaration for hosted weights.
依赖：W2-VLA repository、Transformers、matching robot normalization/config
硬件/传感器：~10.55 GB storage、CUDA GPU
实机证据：Author paper reports bimanual physical evaluations
独立复现：No independent deployment found.
限制/反面证据：No independent execution；Robot-specific calibration and dependencies remain
当前建议：可做权重级最小复现；The fixed, licensed weight package materially lowers inspection cost.
重点标签：VLA权重已开放、腕部预测、近期热门
核验来源：来源 1
W2-VLA把全局任务语境用于预测未来腕部潜变量，并用W2-CoT只在训练时塑造接口。LIBERO为98.5%，RoboTwin 2.0 easy/hard为60.71/18.21，双臂实机三任务标准/OOD平均70.0/52.22%，动作生成87.43Hz。代码和10.55GB权重可取，但两套数据卡均无许可证，仓库LICENSE还附加上游commit保留义务。