Search for: All records

Advancements in robotics and AI have increased the demand for interactive robots in healthcare and assistive applications. However, ensuring safe and effective physical human-robot interactions (pHRIs) remains challenging due to the complexities of human motor communication and intent recognition. Traditional physics-based models struggle to capture the dynamic nature of human force interactions, limiting robotic adaptability. To address these limitations, neural networks (NNs) have been explored for force-movement intention prediction. While multi-layer perceptron (MLP) networks show potential, they struggle with temporal dependencies and generalization. Long Short-Term Memory (LSTM) networks effectively model sequential dependencies, while Convolutional Neural Networks (CNNs) enhance spatial feature extraction from human force data. Building on these strengths, this study introduces a hybrid LSTM-CNN framework to improve force-movement intention prediction, increasing accuracy from 69% to 86% through effective denoising and advanced architectures. The combined CNN-LSTM network proved particularly effective in handling individualized force-velocity relationships and presents a generalizable model paving the way for more adaptive strategies in robot guidance. These findings highlight the importance of integrating spatial and temporal modeling to enhance robot precision, responsiveness, and human-robot collaboration. Index Terms —- Physical Human-Robot Interaction, Intention Detection, Machine Learning, Long-Short Term Memory (LSTM) 

Abstract Humans can physically interact with other humans adeptly. Some overground interaction tasks, such as guiding a partner across a room, occur without visual and verbal communication, which suggests that the information exchanges occur through sensing movements and forces. To understand the process of motor communication during overground physical interaction, we hypothesized that humans modulate the mechanical properties of their arms for increased awareness and sensitivity to ongoing interaction. For this, we used an overground interactive robot to guide a human partner across one of three randomly chosen paths while occasionally providing force perturbations to measure the arm stiffness. We observed that the arm stiffness was lower at instants when the robot’s upcoming trajectory was unknown compared to instants when it was predicable - the first evidence of arm stiffness modulation for better motor communication during overground physical interaction. 

Many anticipated physical human-robot interaction (pHRI) applications in the near future are overground tasks such as walking assistance. For investigating the biomechanics of human movement during pHRI, this work presents Ophrie, a novel interactive robot dedicated for physical interaction tasks with a human in overground settings. Unique design requirements for pHRI were considered in implementing the one-arm mobile robot, such as the low output impedance and the ability to apply small interaction forces. The robot can measure the human arm stiffness, an important physical quantity that can reveal human biomechanics during overground pHRI, while the human walks alongside the robot. This robot is anticipated to enable novel pHRI experiments and advance our understanding of intuitive and effective overground pHRI. 

This work challenges the common assumption in physical human-robot interaction (pHRI) that the movement intention of a human user can be simply modeled with dynamic equations relating forces to movements, regardless of the user. Studies in physical human-human interaction (pHHI) suggest that interaction forces carry sophisticated information that reveals motor skills and roles in the partnership and even promotes adaptation and motor learning. In this view, simple force-displacement equations often used in pHRI studies may not be sufficient. To test this, this work measured and analyzed the interaction forces (F) between two humans as the leader guided the blindfolded follower on a randomly chosen path. The actual trajectory of the follower was transformed to the velocity commands (V) that would allow a hypothetical robot follower to track the same trajectory. Then, possible analytical relationships between F and V were obtained using neural network training. Results suggest that while F helps predict V, the relationship is not straightforward, that seemingly irrelevant components of F may be important, that force-velocity relationships are unique to each human follower, and that human neural control of movement may affect the prediction of the movement intent. It is suggested that user-specific, stereotype-free controllers may more accurately decode human intent in pHRI.