Your Robot’s Safety Functions Already Work. What If the Input Lies?
Article examines how manipulated or incorrect sensor inputs can compromise robot safety systems, with research showing vision-language-action models vulnerable to adversarial interference.