Study finds a learning robot can only forget one user's corrections in some cases

Eshika Pathak and Leela Krishna have studied whether a robot that keeps learning after deployment can forget one of the people who corrected it, in a paper accepted at the HumanRLI workshop at CoRL 2026.

In two simulated tasks with scripted users, they found that when users correct the robot by taking over control, retraining on the remaining sessions in their original order gets as close to the target robot as training noise allows. Simply deleting that person's data and retraining once does not work, because the order of training is lost.

When users only nudge the robot's actions, the other users' corrections carry the removed person's influence, and none of the tested methods recovered the target robot, the authors report.