Introduction

Robot safety systems are designed to stop dangerous actions, but they depend on the very information that adversaries can manipulate. When a robot's sensors, estimates, or messages are altered, the machine may act exactly as programmed—yet cause harm because the input was never what it seemed.

What Happened

Research from VicOne LAB R7 shows that adversarial images, hidden text, and inaudible audio can change robot behavior without changing the assigned task. In one test, text on a poster was treated as an instruction and redirected a robot dog. In another, crafted audio in a hospital-simulation robot altered its simulated actions while the task remained unchanged. These demonstrations reveal that a robot can follow its rules while operating on false data.

Why This Matters

Safety assessments often assume faults are accidental and rare. Deliberate cyber manipulation changes that equation. An attacker can time and repeat a false input, turning a rare glitch into a repeatable threat. Redundant sensors may also share a single point of failure if one manipulation affects multiple inputs, reducing the expected safety margin. Standards such as China's GB/T 45502-2025, IEC TS 63074, and the upcoming EU Machinery Regulation require protection of safety-relevant data, but evidence must reflect real-world manipulation scenarios.

Key Takeaways

  • Safety functions that rely on untrusted inputs can be bypassed without changing the robot's code.
  • Adversarial attacks can target vision, audio, and inertial sensors simultaneously.
  • Redundant systems offer less protection if a single manipulation compromises several sensors.
  • Security standards and lifecycle monitoring are essential to close the gap between safety and cybersecurity.

Conclusion

Robot safety cannot be evaluated in a vacuum. When the information safety decisions rely on is manipulated, protective functions may fail silently. Engineers must map every protective input, test it under adversarial conditions, and agree on response policies before deployment. Only then can robots operate safely in the real world.