Robot alarms can predict failure yet still raise task costs

When a robot monitor raises an alarm, should the system stop, try to recover, or keep going? A signal that identifies likely failures may still be a poor basis for action if it arrives after recovery is possible or disrupts a run that would have succeeded.

It remains unclear whether acting on an alarm improves task completion once false alarms and intervention costs count.

RESCUE replays saved robot-policy states under alternative actions, then weighs their chances of success against the expected cost of acting on monitor alarms.

RESCUE restores saved robot-policy states and tests alternative actions using the original attempt’s remaining budget. It estimates whether each action could finish the task and its expected cost, including resets and retries. This separates whether recovery remains possible from whether acting is preferable.

In the primary OpenVLA-OFT/LIBERO-Plus cell, steps-since-progress had a 7.25% held-out false-alarm rate against a 1.01% break-even rate; its alarms raised cost by 0.0835 reset-equivalents per completed task versus never intervening.

Across the study’s monitor verdicts, detection and timing criteria often admitted monitors, but decision value admitted none against that baseline.

These findings come from simulations in MuJoCo on LIBERO and LIBERO-Plus, where all observed failures were timeouts.

The study tested two robot policies and only Continue and one fixed scripted recovery; its declared branch perturbation and cost model do not establish results on hardware, unseen tasks, or with richer recovery options.