Natural-language instructions given to service robots are often ambiguous, especially when a scene contains multiple possible objects or actions. A new paper on arXiv proposes an adaptive question generation mechanism that lets a robot ask a clarifying question when it is uncertain about what the user meant.

The mechanism is based on two ideas: information gain, to choose the question that most reduces uncertainty, and human cognitive characteristics, to make the question easy for a person to answer. The authors suggest this approach could improve how robots handle incomplete or vague commands in real-world settings.

The paper is at the proposal stage, and the abstract does not include experimental results. Still, the approach points toward more interactive and human-aware robot instruction following.