Robot safety functions can fail when manipulated inputs mislead them
AIVicOne LAB R7 tested how untrusted inputs can change robot behavior, including a robot dog that followed text on a poster and a hospital-service robot simulation altered by crafted inaudible audio. In one bounty test, an injected ROS 2/DDS message made a robot move under a setting organizers expected to keep it still.
Source details
IEEE Spectrum · AI · VicOne · spectrum.ieee.orgPublished · added here
Original title: Your Robot’s Safety Functions Already Work. What If the Input Lies?