Researchers taught a robot to insert a tightly fitting computer connector more reliably by limiting when lower-quality demonstrations influence its training. On one robot platform, the method raised success from 48.3% with precise examples of the task alone to 80% when those examples were supplemented by two imperfect data sources.
Precision assembly usually calls for carefully guided demonstrations of the exact task, which take time to collect. The team tested whether two less suitable sources could help: less accurate demonstrations of the same connector insertion, and accurate demonstrations of a different plug-insertion task. Simply mixing both sources into training throughout the process reduced success on the connector task to 38.3%.
Training uses each imperfect source where it helps
The researchers' method, called ε4P, trains a robot action model with demonstrations at different levels of noise added to the action sequence. When an action sequence is highly noisy, the model relies more on task context, so less accurate examples of the target task can still show what needs to be done. When the action sequence is nearly clean, movement detail matters more, allowing precise examples from the other task to contribute without imposing as much of their mismatched context. The method assigns each source to the part of training where it is useful; it does not change the model architecture or training objective.
Selected imperfect data improves insertion and approaches added precise data
The team trained policies for a Rainbow Robotics RB-Y1 robot and tested insertion of a 24-pin ATX connector into a motherboard, a two-stage cable-plugging task and bolt-and-nut sorting. Success rates reflect 60 real-robot trials per tested policy and task. On the ATX task, using both imperfect sources in their selected training regions produced 80% success, compared with 48.3% for precise target-task data alone and 38.3% when all three sources were mixed throughout training. The paper also reports 88.3% full-task success on cable plugging, versus 78.3% with precise target-task data alone.
A separate comparison tested whether imperfect examples could stand in for an equal number of additional precise examples. On ATX insertion, adding 100 lower-precision demonstrations of that task in the selected training region yielded 73.3% success; adding 100 more precise demonstrations yielded 75%. Those results concern success in the authors' controlled trials, not the cost of collecting data. The experiments used one robot platform and a small set of tasks, so they do not yet establish that the method transfers to other robots or production lines.
Sources
- Wei and colleagues' research paper (posted September 22, 2026).
- Full paper, methods and real-robot results.
- Image: ATX motherboard connectors by Tobias Maier, CC0. Illustrative component, not the hardware used in the study.
