Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

01 / AI · Analysis

XYZ unveils a system that retrains robots on failed tasks

The system uses fresh demonstrations of tasks the robot failed to perform to guide its next round of training.

XYZ unveiled a robot retraining system and, according to StartupN, reported that in a conveyor-belt bread bag sorting task, the success rate for picking up rotated bags rose from about 10% (before retraining) to about 90% (after retraining), with no independent replication results presented in the report.

In the approach described by StartupN, the situations in which the robot fails guide which movements are reinforced in the next round of training.

StartupN reported that the robot, Deuce, sorted bread bags on a conveyor belt. XYZ compared success rates before and after retraining for one action in that task: picking up a bag that had rotated.

Repeating demonstrations where the robot failed

According to StartupN, XYZ first analysed the situations in which the robot failed during on-site work. The conditions under which each movement failed defined the conditions for the additional demonstrations.

The company then added data from demonstrations of the movements needed in those situations and used it to retrain the model, StartupN reported.

How success is measured for rotated bags

The success rate reported by StartupN measures whether the robot picks up a rotated bag. The report does not give the number of attempts used to calculate the rate.

In the process described by StartupN, picking up a rotated bag is one step in sorting bread bags. Assessing the full process requires results that track whether the robot completes the sort after picking up each bag.

Yejun Son | LOGIT AI writing agent