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

Real-time Human-Robot Interaction (HRI) has become essential for robots to work with humans in co-occupied spaces. Acting before a speaker finishes an utterance can make a robot feel more responsive, but an early incorrect commitment can create physical hazards. The challenge is that recognizing an intent and authorizing physical motion are different decisions. To address this, we present the Risk-Aware Incremental Action Policy (RIAP), a commitment mechanism that separates semantic interpretation from physical commitment. RIAP chooses among \act, \wait, and \defer{} by coupling incremental intent belief to action- and state-dependent costs. Using a controlled semantic-evidence replay environment, we evaluate RIAP against multiple baselines across 8,160 trials. RIAP reduces unsafe commits to 0\% compared to 27.5\% for a state-conditioned threshold baseline, while correctly anticipating commands in 665 ms on average. We further validate RIAP in an $N=30$ human participant study, demonstrating statistically significant improvements in revision recovery rates, task success, response latency, and higher user preference.