The AI challenge hidden in everyday life: making sense of messy signals

Artificial intelligence is often judged by how well it performs on clean, structured data. But the real world is a busy place - full of noise, movement, and changing conditions. A hand gesture may barely ripple a Wi-Fi signal. A wearable heart sensor may capture both heartbeat and body motion. A face image may be blurred or poorly lit. Clinical records may reflect differences in age and gender, while crop fields shift with weather, soil, water, and disease pressure. The challenge is whether AI can still find useful meaning in all this messiness.

A series of studies from the Asian Institute of Technology (AIT) points to a shared research direction: building AI that can still make useful decisions when signals are noisy, changing, or incomplete. Across AIT’s engineering, ICT, health-data, and agricultural-technology research, the common goal is to make AI more dependable in real-world environments. This direction can be described as robust sensing intelligence, AI that can still recognize patterns and support decisions when data are imperfect. Rather than waiting for ideal information, these systems aim to extract meaning from the signals already present in people's daily lives.

A particularly visible example comes from AIT’s work on Wi-Fi-based gesture recognition. Instead of cameras or wearable devices, the system analyzes Wi-Fi Channel State Information (CSI) to detect subtle changes in wireless signals caused by hand movement. Using low-cost Raspberry Pi hardware, the Nexmon tool, preprocessing, and a convolutional neural network, the study demonstrated strong performance across different environments, including 97% accuracy in line-of-sight conditions, 93.3% in non-line-of-sight conditions, and 87.9% when both scenarios were combined. The significance goes beyond gesture control. It shows how ordinary infrastructure, such as Wi-Fi already present in homes and buildings, can become a privacy-preserving sensing layer. This could support smart homes, healthcare monitoring, eldercare, and virtual-reality interaction without relying on cameras or body-worn devices.

“Many researchers around the world are working on AI for sensing, healthcare, wireless communication, identity verification, and agriculture,” said  Prof. Attaphongse Taparugssanagorn of the Department of Computer Science and Communication Technology at AIT. “What is important here is the broader connection. These studies show how AI can be designed to work with imperfect real-world data, not only in one field but across many areas that affect people and society.”

The same principle appears across AIT’s broader research. In wearable healthcare, AI is combined with signal-processing methods to improve arrhythmia detection from ECG signals, where motion and sensor variability can reduce reliability. In wireless communication, graph-based and temporal deep learning help recognize modulation patterns under low signal-to-noise conditions, supporting more reliable connected systems. In digital identity, Transformer-based aggregation helps strengthen face verification when images are affected by pose, lighting, occlusion, or low quality.

The approach also extends to public health and agriculture. In diabetes prediction, ensemble learning classifies people as diabetic, non-diabetic, or pre-diabetic while accounting for age and gender, a step toward earlier, more personalized risk assessment. In precision farming, UAV imagery, IoT environmental data, vegetation indices, and machine learning are integrated to predict maize yield earlier, helping farmers make better crop-management decisions.

The individual fields are not new in themselves. Researchers worldwide are actively studying AI for ECG analysis, wireless systems, gesture recognition, face verification, diabetes prediction, and crop-yield forecasting. The distinctive value of the AIT work is its integrated perspective: these studies share a common challenge, making AI useful when data come from real-world environments rather than controlled laboratories.

The researchers note that broader validation is still needed across diverse users, devices, populations, environments, crops, and deployment conditions. For health and identity applications, privacy, fairness, and responsible use will also be essential. In other words, the next generation of useful AI may not be the one that waits for perfect data, but the one that helps people act wisely with the imperfect signals already available to them.