Your Robot’s Safety Functions Already Work. What If the Input Lies?

Robots can be misled by false or malicious inputs, causing unsafe actions. Researchers at VicOne LAB R7 show that visual patches, audio cues, and text prompts can trick vision‑language‑action models like Gemma 4 E4B and Nemotron on Jetson AGX Orin into ignoring safety signals. This highlights the need for robust input validation.

Opening Kapyn…