- 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:released for critic training/evaluation
权重:released separately
数据:released separately
许可证:MIT;Repository code only.
依赖:Python 3.12、uv、PyTorch、LeRobot v3、ViT、CLIP、ffmpeg
硬件/传感器:single CUDA GPU for quick start、paper uses A100 and WidowX-250S
实机证据:Seven WidowX tasks; paper test table uses 50 trials per method/task
独立复现:None found.
限制/反面证据:Full VLA-RL integration is not completely released;No independent reproduction
当前建议:高潜待持续验证;A runnable critic loop and physical data make the idea inspectable, but the complete RL stack remains partial.
重点标签:VLA强化学习主流技术2.0、世界模型critic、可能成为新主流、近期热门
核验来源:来源 1
- 数据 · 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:released separately
数据:181 real-robot episodes released
许可证:MIT;Dataset-card declaration for this hosted package.
依赖:LeRobot v3、datasets/pyarrow、video decoder
硬件/传感器:Storage for released episodes
实机证据:181 WidowX pick-and-place episodes
独立复现:No independent training report found.
限制/反面证据:Only one released real task subset;Not the full seven-task paper corpus
当前建议:可做最小复现;The small released subset supports a cheap critic-only verification.
重点标签:VLA critic数据、实机小样本、近期热门
核验来源:来源 1
- 权重 · 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 separately
权重:best.pt downloadable
数据:released separately
许可证:not_disclosed;No model-card license; code MIT is not automatically a weight license.
依赖:WCM code and matching config
硬件/传感器:~2.2 GB storage、CUDA GPU
实机证据:Author-declared pick-and-place checkpoint
独立复现:None found.
限制/反面证据:Weight license absent;No independent checksum execution
当前建议:许可证待澄清;Downloadability is verified, but legal reuse remains unresolved.
重点标签:权重已上传、许可证缺失、重点更新跟踪
核验来源:来源 1
WCM把动作条件潜空间世界预测加入VLA强化学习critic,在149个仿真任务和WidowX-250S七项实机任务上验证。论文实机每任务100条SFT轨迹、8轮RL且每轮50条rollout,最终表格按每方法50次测试;固定MIT仓库已有可运行critic训练评测,公开181-episode数据与约2.2GB checkpoint,但完整VLA-RL代码和论文全部检查点仍承诺逐步开放。