ALERT: Anticipating Learner Errors for Robot Teaching
A. S. Cheng, J. Lin, X. Sun, N. Walker, J. X. Liu, and A. Bobu, “ALERT: Anticipating Learner Errors for Robot Teaching,” in CoRL 2026 Workshop on Human-Centered Robot Learning and Interaction, Nov. 2026.
Abstract
Training generalist robot policies takes vast amounts of demonstration data, and existing teleoperation systems demand continuous human attention and control to collect it. To reduce the cognitive and physical load on operators, we propose ALERT, a data collection system that combines the base policy being fine-tuned, a proactive intervention detector, and an augmented-reality (AR) visualization. While the base policy drives the robot, the detector anticipates the policy’s mistakes before they occur and asks the operator to intervene only when needed. The AR view overlays the policy’s next actions onto the robot, allowing the impending failure to be visible so the operator can intervene correctly. Trained across diverse tasks, the detector generalizes to held-out ones with 68% accuracy and 0.69 F1 in simulation and 60% accuracy and 0.68 F1 on a real robot. In a user study, operators performed better on a secondary task and reported lower cognitive and physical load with ALERT than with the baselines. By prompting operators to intervene only when needed instead of asking them for full demonstrations, ALERT opens the way to multi-robot data collection with a single operator, easing the data bottleneck that limits generalist policies.
BibTeX Entry
@inproceedings{cheng2026alert,
author = {Cheng, Amanda Sijia and Lin, Jeff and Sun, Xiangzhou and Walker, Nick and Liu, Jason Xinyu and Bobu, Andreea},
title = {ALERT: Anticipating Learner Errors for Robot Teaching},
year = {2026},
month = nov,
location = {Austin, TX, USA},
booktitle = {CoRL 2026 Workshop on Human-Centered Robot Learning and Interaction},
}