Event

Samsung proposes ε4P: training robot policies with imperfect data

1 reports · 1 sources

The Samsung RX Lab-China team proposed the ε4P (Imperfection for Precision) method, which admits imperfect data into different noise intervals of flow matching according to the type of data defect, turning data that is harmful when mixed directly into a performance gain. The method only adjusts the noise sampling ranges of different data sources: high-precision data for the target task covers the full interval, low-precision target-task data is used for the high-noise interval, and high-precision data from other tasks is used for the low-noise interval; the model architecture, the flow-matching objective, and the data-source sampling probabilities all remain unchanged. The team tested on three real-robot tasks—ATX 24-pin connector insertion, two-stage cable plugging, and bolt-and-nut sorting—and reports that across the two types of imperfect data and the three tasks, the average success rate dropped by only 4.2 percentage points, which is expected to reduce reliance on high-quality data.