A humanoid robot decides when to speak up—and when to hold back

How the robot chooses to act

ProAct combines two parts. A Behavioral System handles low-latency speech and movement, including a motion generator that can add a new intention while a continuous movement stream is already running. A Cognitive System keeps track of dialogue and visual context, predicts whether a user may be motivated to receive help or engagement, and sends a high-level intention to the Behavioral System.

In five interaction scenarios, the robot generated its own speech and movements rather than following prescribed body-motion instructions. In one example, it turned toward a passerby, guided the conversation to a poster and used a synchronized gesture while correcting a verbal misconception. In another, it noticed that a user was looking at a smartphone and prompted them to re-engage, again using a gesture. The evaluation protocol measured whether the system triggered on positive cases while remaining quiet on negative ones, along with response latency, human ratings and failure rates. The supplied excerpts do not report the resulting scores.

Why timing matters

A physically embodied social robot needs to do two things at once: respond quickly enough to remain part of a conversation and reason over longer-term context before interrupting or offering help. ProAct’s split design is intended to let slower social reasoning run without stopping the robot’s immediate speech and movement, giving proactive behavior a way to enter an ongoing interaction.

Evidence and open limits

The work was evaluated through real-world user studies, motion-generation benchmarks, qualitative interaction examples and ProActBench, which contains 40 cases covering major types of proactive triggers. The benchmark used one environment and one decision point per case; its 10 negative-trap cases limit analysis of false alarms to paired comparisons. Replaying the same inputs across systems helps isolate decision-making, but it cannot show how people adapt to a robot’s behavior during a live interaction.

The authors also report that triggering uses fixed intervals and thresholds, which may miss brief opportunities or cause unnecessary interventions. ProAct’s first verbal response begins about two seconds after the user finishes speaking, partly because it relies on a cloud-based language model. The system mainly uses posture, facial expression, dialogue and scene history, so the examples involving emotional support do not establish reliable recognition of subtle feelings in unrestricted settings.