When Robots Mishear Us: Mapping the Safety Risks of Voice-Controlled Embodied AI

Voice commands create a new safety vulnerability: When voice-controlled robots mishear user instructions, they can execute harmful actions that the model's safety systems would normally block. This represents a distinct safety risk beyond what text-based AI systems face.

Some mishearings are particularly dangerous: The research found that certain types of speech recognition errors preserve the general meaning of a command while introducing harmful ambiguity—making it harder for the AI to recognize that an instruction is unsafe. For example, a mishearing might transform a benign request into one that bypasses safety guardrails.

Current safety measures are incomplete: Existing safety benchmarks (SafeAgentBench and POEX) don't adequately account for speech recognition errors, meaning robots tested as "safe" in controlled conditions may fail in real-world voice interactions where mishearings occur naturally.

Automatic error correction isn't a reliable fix: While the researchers tested whether automatically correcting speech recognition mistakes could mitigate risks, they found this approach is inconsistent and doesn't reliably prevent unsafe outputs—suggesting we need more robust solutions.

This gap matters now: As voice-controlled robots become more common in homes and workplaces, the combination of speech recognition errors and embodied AI systems that can physically act creates urgent safety challenges that current development practices don't adequately address.