EGM Curated G1 Motion Corpus
EGM论文从AMASS与LAFAN1重定向到23-DoF Unitree G1，人工筛出674条/4.08小时高质量训练动作，并构建13,596条/49.25小时未见测试集；另有10,393条/27.47小时规则与策略过滤大集。作者报告小集优于大集，但没有发布序列清单、重定向结果、split、失败动作、处理代码、许可证映射或下载入口。
- 开放状态
- 项目已核验·制品未开放
- 许可证
- No EGM dataset license or derivative-rights ledger; AMASS and LAFAN1 retain upstream terms
- 商业使用
- not established and subject to upstream source-dataset terms
- 人工评审结论
- paper_claimed_amass_lafan1_derived_g1_motion_corpus_not_downloadable
- 机器人
- Unitree G1 23-DoF simulation representation
- 任务
- general motion tracking、high-dynamic motion tracking、teacher policy training、DAgger student distillation
- 模态
- retargeted G1 joint motion、root pose、joint velocity、motion category、tracking-error-derived sampling bins
- 格式
- not released
- 采集/生成方式
- Isaac Gym
- 已知问题
- No files/card/split/checksum/repository/license; author-defined manual quality screen; source derivative rights unresolved; comparison uses author reimplementation of GMT; no physical result or independent audit.