Toward Human-Like Assistance: Detecting Help-Seeking in Human�Robot Collaboration Via Implicit Signals
Ane San Martin∗, Ander Iriondo, Michael Hagenow, Julie Shah, Johan Kildal, and Elena Lazkano
AI summary
Problem
Detecting when a human collaborator needs help in industrial settings typically relies on explicit requests, which are often delayed or absent, limiting proactive robot assistance.
Approach
The authors trained machine learning models on implicit nonverbal signals (facial expressions, gaze, head pose, and affective states) collected during a collaborative assembly task, using temporal sliding windows to capture behavioral dynamics.
Key results
- Random Forest achieved an F1-score of 0.98 with averaged temporal features
- Affective states (valence and arousal) were the strongest predictive features
- Larger temporal windows consistently improved model accuracy
- A 3–4 feature subset maintained high accuracy for efficient deployment
Why it matters
Enables industrial cobots to proactively assist workers by interpreting subtle nonverbal cues, reducing task delays and cognitive load in human-robot collaboration.
Abstract
As collaborative robots become more common in industrial settings, enabling them to provide timely support to people when they are stuck is a key aspect. In particular, detecting when users need assistance, without relying on explicit requests, remains an open problem. This work explores whether machine learning models, such as Random Forest and Decision Tree, as well as the use of temporal dependencies can detect help-seeking behavior from implicit signals. Through a user study in a robot-assisted assembly task, we show that nonverbal cues, such as affective states and subtle behavioral dynamics, can reliably predict when a human needs assistance. Our top- performing model achieved an F1-score of 0.98. These findings demonstrate the feasibility of leveraging temporal modeling of implicit signals for proactive interaction in industrial contexts.