Not by chance: exploiting flaws in humanoid robot safety filters

Researchers propose a goal insertion attack against humanoid robots and develop an automated tool, SIREN, showing that existing safety filters can themselves become a source of safety risk. Safety filters use mathematical guarantees to ensure that control actions will not cause collisions, but this guarantee holds only when the safety parameters are configured correctly, and correct configuration varies by task and environment; there is currently no general method to determine it, so the safety filters shipped with humanoid robots may fail in specific scenarios after deployment. An attacker needs only ordinary user interaction, with no additional privileges, to exploit these failures. In experiments, SIREN identified attacks in all vulnerable scenarios in the SPARK dataset, found three times as many attacks as baseline methods, and caused a real humanoid robot to collide while executing tasks.