Advanced robots rely on continuous data inputs to perceive their surroundings and execute tasks. However, research from VicOne LAB R7 and other published studies suggests that manipulating these inputs could dangerously alter robot behavior. In a simulated test of vision-language-action (VLA) models, introducing a visual patch into the camera's view reduced task success. Similarly, during a test involving a robot dog powered by Gemma 4 E4B, text displayed on a poster was processed as a direct instruction, which changed the machine's movement patterns.
Intentional data manipulation may also bypass independent protective mechanisms. For instance, if a humanoid robot's balance controller depends on a vulnerable gyroscope, acoustic waves at specific resonant frequencies could distort its rotation data. While researchers have not demonstrated this exact attack causing a humanoid to fall, the underlying acoustic vulnerability has been proven in drones. Furthermore, injecting a ROS 2/DDS message into a robot caused it to move when it was expected to remain completely stationary under safe-control settings.
Safety teams evaluating physical risks in robotics must account for deliberate cyber threats. Redundant safety measures or sensor fusion techniques might fail to provide protection if they rely on the same manipulated data source. Experts advise that before deployment, robots should be thoroughly tested against both accidental faults and adversarial scenarios to ensure their safety limits remain effective under manipulation.