TrajBooster AgiBot-to-Unitree G1 Retarget
TrajBooster把AgiBot World视觉与语言轨迹中的执行器动作替换为Unitree G1全身重定向动作,官方口径为176个任务、1,960个episode、约35小时;固定Hugging Face修订公开非gated,共9个文件、约29.28GB。数据卡声明Apache-2.0,但源AgiBot World为CC BY-NC-SA 4.0,派生仓许可不能自动清除上游非商用和视觉/语言/资产权利。
- 开放状态
- 项目制品开放·上游条款另适用
- 许可证
- 派生数据卡Apache-2.0;源AgiBot World仓声明CC BY-NC-SA 4.0;源视觉、语言与资产条款需单独核验
- 商业使用
- not established because source dataset terms are non-commercial
- 人工评审结论
- public_29_28gb_1960_episode_cross_embodiment_derivative_with_upstream_rights_conflict
- 机器人
- source AgiBot embodiments、target simulated Unitree G1 humanoid with Dex-3-oriented action space
- 任务
- 176
- 模态
- multi-view RGB retained from source、language instructions retained from source、source observations、retargeted Unitree G1 whole-body actions、LeRobot/Parquet episode metadata
- 格式
- LeRobot/Parquet-style dataset split across seven large chunks plus metadata files
- 采集/生成方式
- retargeting and simulated interaction pipeline described by TrajBooster; exact end-to-end raw-source regeneration closure is incomplete
- 已知问题
- 目标动作由重定向生成而非真机采集;派生Apache元数据与源AgiBot World非商用条款存在权利边界;没有完整raw-source再生成闭包;没有失败轨迹比例、episode-to-real配对或独立VLA指标复算。
