PACT: Robots that ask before acting improve long-term collaboration
New AI framework lets robots know when to ask for help instead of guessing wrong.
Deep Dive
The paper introduces PACT (Proactive Asking for Continual Task Assistance), an ask-or-act framework for human-robot collaboration. It uses reinforcement learning and cross-day interaction history to evaluate contextual sufficiency before acting, and introduces a clarification utility metric to balance assistance accuracy and asking frequency. In multi-day experiments, PACT consistently improves both assistance accuracy and clarification utility compared to passive inference baselines.
Key Points
- PACT uses reinforcement learning to decide when to ask for clarification versus when to act, leveraging cross-day interaction history.
- It introduces a clarification utility metric that quantifies the trade-off between assistance accuracy and the frequency of user interruptions.
- In multi-day experiments, PACT outperformed passive inference baselines in both assistance accuracy and overall utility.
Why It Matters
Enables robots to learn user preferences over time, making long-term home or office assistance more reliable and less frustrating.