RIAP: Risk-Aware Incremental Action Policy for Physical Commitment from Partial Speech

While acting early on conversational cues makes a robot feel responsive, premature movement can introduce physical hazards. The root challenge is that recognizing a human's intent and authorizing robotic motion are separate decisions. We solve this with the Risk-Aware Incremental Action Policy (RIAP), a framework that decouples these decisions by weighing real-time intent confidence against action- and state-dependent safety costs. RIAP continually selects the optimal behavior from three choices: ACT, WAIT, or DEFER. Through 8,160 simulation replays, RIAP cut unsafe commitments from 27.5\% to 1.6\%, anticipated user commands by an average of 1,022 ms, and corrected for 60.6\% of mid-utterance human revisions—a trade-off profile that existing decision mechanisms could not reproduce. A 30-participant study further confirms that RIAP delivers a significantly superior perceived user experience across all measured subjective dimensions.