Humanoid robots can fall and damage themselves. Existing safety monitors only learn from states the robot has already visited, so in unfamiliar states they tend to assume recovery is possible. That is dangerously optimistic.
The authors learn a recoverability envelope: the set of states from which a fixed fallback policy can still avoid falling. Their framework alternates between learning this envelope and actively exploring, seeking unknown states without sampling clearly unrecoverable ones. States with high uncertainty are treated as unrecoverable.
On stochastic benchmarks and a real humanoid, this covered far more of the true envelope and predicted failures more accurately than prior methods, transferring to hardware without retraining. That makes runtime fall monitors safer to trust.