FailBench lets home robots fail gracefully with new safety model

Service robots entering homes face countless unpredictable hazards — a cat knocking over a vase, a child grabbing an arm, or a sensor suddenly glitching. While roboticists try to prevent failures, some are unavoidable. A new paper from researchers led by Duc Nguyen tackles this head-on by reframing the problem: instead of only avoiding failure, robots should plan for how to fail gracefully. The team introduces a safety formulation that calculates both the probability of impactful interactions between a robot and nearby entities (humans, pets, objects) during a failure, and the severity of those outcomes. This two-part metric lets a robot make informed motion-planning decisions that weigh safety against task efficiency, so it can continue operating where possible while minimizing harm when things go wrong.

To support systematic evaluation, the paper also presents FailBench, a MuJoCo-based simulation framework built specifically for studying robot-environment interactions under diverse failure modes such as sensing degradation and actuator malfunctions. FailBench provides a standardized way to test how different safety-aware planners behave when failures are injected mid-task, filling a gap in existing robotics benchmarks. The work has been accepted to IEEE ICRA 2026 and is available on arXiv (2608.05313). For developers building service robots, this research offers a path toward safer deployment: quantifying failure impact in advance rather than reacting after the fact. It also suggests a shift in how we judge robot reliability — from perfect operation to resilient, safe degradation in unpredictable human spaces.