Introduction
The rapid advancement of Artificial Intelligence (AI) has transformed fields such as computer vision and autonomous robotics. However, a critical frontier remains: Human-Centred Artificial Intelligence (HAI). While traditional AI systems often operate in isolation or follow rigid, pre-defined scripts, the next generation of intelligent systems must be capable of understanding, predicting, and collaborating with human movement. This editorial explores a collection of research that bridges the gap between raw data and human intention, demonstrating how AI can move beyond simple automation to become a nuanced partner in industrial, clinical, and creative settings.
The Challenge of Human-Centred AI
The integration of AI into daily life—whether as a collaborative robot (cobot) on a factory floor or a health-monitoring wearable—requires machines to "read" human intentions through physical movement. The authors of this Research Topic identify two primary hurdles:
- Sensing Accessibility: Transitioning from expensive, lab-grade Motion Capture (Mocap) systems to "modality-simple" and cost-effective sensors, such as smartwatches or basic cameras, that can be used in real-world environments.
- Algorithmic Interpretation: Developing models capable of handling the temporal complexity of human motion. It is no longer enough to recognize what a person is doing; systems must understand how and why they are doing it to provide meaningful assistance.
Methodological Innovations in Movement Analysis
The research synthesized in this editorial highlights several novel approaches to movement sensing and interpretation:
Force Myography (FMG) for Intuitive Control
One study explores the use of Force Myography—employing force-sensing resistors (FSRs) on the arm—to estimate isometric force and torque. By utilizing a two-stage regression strategy with models such as General Regression Neural Networks (GRNN), Support Vector Regression (SVR), and Random Forest (RF), researchers can create more intuitive human-robot interfaces that respond to muscle pressure rather than just joint position.
Modeling Professional Gestures
To assist in industrial and craft settings, a multivariate time series approach using State-Space modeling was introduced. This method recognizes professional gestures and forecasts trajectories by analyzing the dynamic associations between different body parts, allowing for more fluid collaboration between human workers and automated systems.
Real-Time Clinical Monitoring
In the healthcare domain, the "DeepFoG" architecture was developed to detect Freezing-of-Gait (FoG) in Parkinson’s patients. Remarkably, this deep learning model achieves real-time detection using data from a single wrist-worn Inertial Measurement Unit (IMU), offering a non-intrusive alternative to complex multi-sensor arrays.
Gamification and Disease Progression
The "i-Prognosis Game Suit" represents a shift toward "invisible" monitoring. By translating clinical motor assessment standards (such as the UPDRS Part III) into digital games, researchers can monitor disease progression through play, providing a continuous stream of data that is less burdensome for patients than traditional clinical visits.
Key Insights: From Recognition to Collaboration
A recurring theme across these works is the power of single modalities. The research demonstrates that a single, low-cost sensor can achieve high accuracy if paired with a sufficiently sophisticated deep learning model. This shift is crucial for the economic and societal scalability of HAI.
Furthermore, the research moves from simple activity recognition (e.g., "walking") to understanding motor learning and adaptation. By analyzing how humans adapt parameters in probabilistic models or plan movements via reinforcement learning, AI can better anticipate human needs. This transition is essential for addressing the societal challenges of HAI, including safety, transparency, and ergonomics.
Results and Future Directions
The results presented in this Research Topic are promising. The DeepFoG model successfully differentiated between "stops" and "walking with turns" using only wrist data, and the i-Prognosis games were validated against gold-standard clinical scales. Additionally, the FMG-based force estimation proved feasible for complex 3-DoF and 6-DoF tasks.
However, significant work remains. Integrating these individual components into a cohesive "perception layer" for autonomous entities like drones or IoT devices is still in its early stages. Future research must focus on:
- Real-World Generalization: Moving models out of controlled environments and into unpredictable, "in-the-wild" settings.
- Predictive Forecasting: Enhancing the ability of AI to forecast movement trajectories before they are completed, enabling proactive rather than reactive assistance.
- Optimal Multi-modal Fusion: Identifying the "sweet spot" where adding a secondary sensor (e.g., combining RGB cameras with IMUs) provides a significant boost in reliability without compromising cost-effectiveness.
By focusing on the human at the center of the movement, these advancements pave the way for AI systems that are not just tools, but true collaborators in the industries and creations of the future.
Summary updated