ROBOGATE: finding where a robot policy fails before you deploy it
A two-stage boundary-focused sampling method for discovering failure modes in robot policies prior to deployment. Rather than testing average performance, it searches for the edge of the region where the policy works.
Average-case evaluation is nearly useless for physical systems. A policy that succeeds 95% of the time is not 95% safe — it is a system whose 5% needs to be characterised, because that is where the hardware damage and the injuries live.
Boundary-focused sampling attacks that directly by spending the test budget near the failure edge instead of uniformly across the task space. It is the same instinct as adversarial testing in software, applied where failures cost more than a stack trace.
It also connects to the serving and deployment work appearing elsewhere in the literature. The research questions in robotics are shifting from capability ceilings to deployment conditions, which is what a field looks like shortly before it becomes an industry.
arXiv — ROBOGATE: adaptive failure discovery for safe robot policy deployment via two-stage boundary-focused sampling → · arXiv — ROSA: a robotics foundation model serving system for robot factories → · arXiv — Safe-Night VLA: thermal-perceptive vision-language-action models for safety-critical manipulation →