Imperfection for Precision: Upcycling Imperfect Data for High-Precision Robotic Manipulation
别急着丢掉「坏数据」：三星团队提出 ε4P，让不完美数据助力高精度机器人操作
Samsung Robotics eXperience Lab-China proposes ε4P, a method that enables high-precision robotic manipulation by utilizing imperfect data through controlled noise interval assignment during training. The approach identifies and separates data defects, assigning different noise intervals to different data sources to maximize their utility while minimizing negative impacts. Experiments on three real-world tasks show significant performance improvements when using ε4P compared to direct mixing of imperfect data.
The paper introduces ε4P, a method that leverages imperfect data for high-precision robotic manipulation by assigning different noise intervals to data sources during training. This approach allows for better utilization of imperfect data without compromising performance, as demonstrated through experiments on three real-world robotic tasks.
Source: Machine Heart — Robotics on WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: Available Artifact link Open source S10
- dataset: not_reported
- code: not_reported
- weights: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
先看结果：同样的不完美数据，换一种用法就能带来提升 团队在 ATX 24 针连接器插入、两阶段线缆插接，以及螺栓螺母分拣三个真机任务上进行了测试，测试顺序为：仅使用目标任务高精度数据，直接混入两类不完美数据，使用 ε4P 混合训练。
Open source S12
如图 3 所示，ε4P 根据数据缺陷来源，将不完美数据分为两类：一类是 任务匹配但精度不足 的数据，如 UMI 数据，保留了正确的任务信息，但动作监督不够精确；另一类是 精度匹配但任务不匹配 的数据，如其他任务的高精度遥操作数据，动作足够精准，但任务上下文存在偏差。
Open source S19
论文标题：Imperfection for Precision: Upcycling Imperfect Data for High-Precision Robotic Manipulation 论文链接：https://arxiv.org/abs/2609.26672 项目主页：https://varepsilon4p.github.io/ 图 1 ε4P 数据训练示意图。
Open source S10
Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com