CuraEval: Evaluating Demonstration Curation on Mixed-Skill Robot Data

Sanity check for demo-curation methods on mixed-skill robomimic data: one score ranks demos almost exactly like episode length, and after Holm correction no curated subset significantly beats training on all data.对示范数据筛选方法做严格对照：在混合技能的 robomimic 数据上与等量随机子集和全量训练比较，发现某评分几乎等同于按轨迹长度排序；经多重比较校正后，没有任何筛选条件显著优于全量数据训练（CoRL 2026 Workshop 论文）。

@misc{xu2026curaeval,
  title = {{CuraEval: Evaluating Demonstration Curation on Mixed-Skill Robot Data}},
  author = {Jiuyi Xu and Meida Chen and Yang Ye and Yangming Shi},
  year = {2026},
  eprint = {2610.38221},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.38221},
}