Amit Desai

Embodied agents are crossing from answering questions to taking physical actions — moving a box, turning a wheel — and people will command them by voice, because voice is the fastest, most natural interface we have. But voice is also the most error-prone, and when a misheard command drives a physical action, the failure isn't a wrong answer; it's human harm, damage, or an expensive, irreversible mistake. The field has never needed a serious way to handle voice-command errors, because informational agents made them cheap. Embodiment ends that. This talk replaces the usual hand-waving — "don't ask too much, don't get it wrong too much" — with a single number you can optimize. The core idea: both confirming and erring cost the user. A confirmation is friction — attention, time, a delayed action; a wrong action is a mistake cost, often higher given physical harm or expense. Put them on one ledger and you can measure a voice interface as average user cost per command, and make minimizing it the system's objective. From that falls a non-obvious rule — you confirm or not based on both cost and uncertainty: an expected value. I'll frame confirmation as just one option alongside acting, disambiguation (choices), and deferring; reason at the level of goals rather than low-level motion; walk the architecture (task hypotheses → user-cost model → confirmation policy); and show eval results from a simulated environment measuring regret against oracle behavior. I'll close with what worked applying this to voice in smart TVs, speakers, and navigation — and a challenge to bring this metric to robots, cars, and wearables before the errors do.