Abstract
The use of social, anthropomorphic robots to support humans in various industries has been on the rise. During Human-Robot Interaction (HRI), physically interactive non-verbal behaviour is key for more natural interactions. Handshaking is one such natural interaction used commonly in many social contexts. It is one of the first non-verbal interactions which takes place and should, therefore, be part of the repertoire of a social robot. In this paper, we explore the existing state of Human-Robot Handshaking and discuss possible ways forward for such physically interactive behaviours.
Introduction
In social robotics, the "first impression" is often physical. Handshaking is a ubiquitous human greeting that conveys emotion, establishes dominance or warmth, and sets the tone for cooperation. However, for a robot, a handshake is a complex multi-stage problem involving trajectory planning, force feedback, and real-time synchronization with a non-deterministic partner.
This research addresses the lack of a unified framework for human-robot handshaking (HRH). It highlights the gap between purely functional robotic movements and the nuanced, socially-aware interactions required for robots to be accepted in human-centric environments like healthcare or hospitality.
A Unified Framework for Robotic Handshaking
To bridge the gap between mechanical movement and social interaction, researchers have categorized the handshake into a conceptual framework consisting of three technical stages and a final social evaluation phase.
1. Hand Reaching
Early models used simple transfer functions or "minimum jerk" trajectories to mimic human velocity profiles. Modern approaches leverage Machine Learning, including LSTMs to predict a human's final hand pose and Deep Reinforcement Learning to learn physical interactions directly from human-human data.
2. Hand Grasping Control
This stage focuses on the tactile "feel" of the interaction. Techniques range from simple position control to sophisticated impedance control and the use of soft robotics (e.g., silicone-rubber hands). Some systems use "sensor-less" control where a deformable palm provides feedback based on physical displacement.
3. Shaking and Synchrony
To achieve a natural "bounce," researchers use Central Pattern Generators (CPGs)—biologically inspired neural circuits—or Harmonic Oscillator Models (like spring-damper systems). These allow the robot to adapt its frequency to the human partner in real-time, ensuring the interaction feels reactive rather than pre-programmed.
Evaluation and Social Impact
The paper reviews how these technical implementations are tested using metrics like the Model Human Likeness Grade (MHLG) and bipolar scales to measure user perception. These evaluations rely on diverse datasets, including:
- Motion Capture Data: Tracking human joint movements for trajectory models.
- EMG and IMU Data: Measuring muscle contractions and joint acceleration.
- Haptic Feedback Data: Recording pressure and grasping forces.
- User Study Feedback: Qualitative and quantitative data from participants interacting with platforms like Nao, Pepper, or custom robotic arms.
Key Insights and Results
- Synchronization is Key: Low interaction forces during the shaking phase are a primary indicator of successful synchronization. The most successful models are reactive rather than static.
- The "Multi-Modal" Effect: A handshake's quality is significantly improved by a robot’s facial expressions and gaze (e.g., shifting from the hand to the face upon contact).
- Handshakes as Data Sources: Handshakes can be used for differentiation; some studies achieved 75% accuracy in classifying personality or gender based solely on handshake statistics like speed and pressure.
- Trust and Cooperation: Shaking hands with a robot significantly increases a human's willingness to help that robot with a task, proving the social utility of the gesture.
- Performance Benchmarks: Interactive models are closing the gap with human-operated robotic handshakes, scoring around 5.3 to 5.9/10 for likeness compared to 6.8/10 for human-operated and 3.0/10 for non-reactive models.
Limitations and Future Work
Several hurdles remain for the field. There is a notable lack of benchmarking due to the vast differences in hardware and metrics used across studies. Furthermore, current models often assume a Western-style handshake, necessitating future work into cultural nuances and alternative "low-touch" greetings like fist bumps in a post-pandemic context. Finally, the "handoff" between individual phases (reaching to shaking) often feels clunky and requires better integration for a truly seamless experience.
Summary updated