For social robots to be truly integrated into office or domestic environments, they must understand user preferences—such as knowing a specific time for coffee delivery or window adjustments. However, acquiring this knowledge presents a dilemma: passive observation is slow and error-prone, while direct questioning can quickly lead to user annoyance. This research, conducted by a collaborative team from the University of Southampton and the University of Warwick, proposes a novel framework to balance information gain against the "cost" of bothering the user.
The Challenge of Robot Proactivity
Traditionally, robots learn preferences through long-term observation or by asking questions. The "ask" approach is faster but introduces significant user friction. If a robot interrupts a user too frequently, the user may eventually cease interaction. The core challenge addressed in this paper is identifying the "sweet spot": which questions provide the most information for the least amount of bother, and when should the robot stop asking entirely?
A Framework for Annoyance-Aware Elicitation
The researchers formalize this as the User Annoyance Aware Preference Elicitation problem, treating it as a "budgeted" optimization task. Their approach relies on three primary pillars:
- The Preference Tensor: Preferences are modeled as a 3D tensor (Tasks × Time Slots × Users). Assuming human behavior follows patterns, the team uses tensor completion algorithms to predict unknown preferences from a small set of known samples.
- The Annoyance Cost Model: Rather than simply counting questions, the model tracks "Annoyance Cost so Far" (ASF). This accounts for timing; questions asked in rapid succession are more costly than those spaced out over time, allowing the robot's "annoyance budget" to effectively recharge during periods of silence.
- The Knapsack Heuristic: The robot treats question selection like a Knapsack Problem. Each potential question has a "weight" (annoyance cost) and a "value" (reduction in tensor uncertainty). The researchers developed a Multishot Heuristic, an adaptive approach where the robot asks a question, updates its knowledge, recalculates uncertainty, and then selects the next most informative query.
Experimental Evaluation
The framework was evaluated using a high-fidelity Gazebo-based simulation of an office environment featuring a TIAGo Steel mobile manipulator. The simulation included seven common office tasks, such as fetching drinks and emptying bins. To test robustness, the authors created four synthetic user profiles: Random (unpredictable), Neutral (no strong opinions), Specific (high preference for one task/time), and Normal (preferences following a bell curve).
Testing spanned "Small Office" (5–10 users) and "Big Office" (15–20 users) scenarios to ensure the tensor completion approach could scale effectively.
Key Insights and Results
- Efficiency of Multishot Elicitation: The Multishot heuristic significantly outperformed baselines. It recovered user preferences with high accuracy by asking only 3 to 6 questions per user, compared to the 7 to 11 questions required by "Singleshot" or random selection methods.
- Uncertainty as Value: The robot measures value by running three different tensor completion algorithms simultaneously. High disagreement between algorithms indicates high uncertainty, marking that specific preference as a high-priority target for a question.
- Physical Execution vs. Preference: The TIAGo robot successfully completed 73% of its tasks. Interestingly, physical complexity impacted the framework; the robot was highly successful at delivering packages (83%) but struggled with emptying bins (33%). The framework accounts for these failures by lowering the user's preference for a task if the robot fails to execute it correctly.
- Scalability: Error rates remained consistently low even as the user count quadrupled, suggesting that the low-rank tensor assumption holds true for larger social groups.
Limitations and Future Directions
While the mathematical model for annoyance is robust, the researchers acknowledge that an "annoyance budget" is currently a predefined constant. In real-world applications, a user's patience likely fluctuates based on stress levels or the perceived importance of the robot's task.
Future work will focus on validating these findings with real human participants to ensure the mathematical "annoyance cost" aligns with actual human frustration. Additionally, the team aims to refine multi-day optimization to help robots maintain accurate preference maps over long periods without hitting the annoyance ceiling.
Summary updated