When Robots Mishear Us: ASR Errors Pose Safety Risks for Embodied AI Models

Summary

A new arXiv paper (2608.28518) by Sihan Jia and Oliver Lemon investigates whether automatic speech recognition (ASR) errors in user input can cause Embodied AI (EAI) models to produce unsafe outputs. The researchers simulated ASR errors and combined them with established safety benchmarks SafeAgentBench and POEX to evaluate how recognition mistakes affect embodied AI safety. Their findings show that ASR errors can lead to harmful instructions being accepted and executed by EAI models. Some errors preserve semantic structure while introducing harmful ambiguity, while others weaken the model's refusal behaviour, allowing unsafe plans to be generated and executed. The study also examined automatic ASR error correction, showing it can reduce risk in some cases but is not always effective. Overall, the paper demonstrates that ASR errors constitute a significant safety risk for embodied AI systems, highlighting the need for robust speech-to-action pipelines.