Why moving artificial intelligence research from clean labs into the real world remains difficult

Artificial intelligence systems that perform well in controlled lab settings often fall short when used in homes, hospitals, or city streets. University of Mississippi assistant professor Bo Wang highlights three main hurdles in moving AI from the lab to real-world environments: generalisation, human behavior, and resources. Most vision-based AI models are trained on well-lit, high-quality images from curated sources. When deployed in dimly lit rooms or nighttime settings, their performance drops sharply because they struggle to adapt to tougher conditions. Collecting more nighttime training data is difficult and expensive because low light obscures visual information and makes manual labeling hard for human annotators. To bridge this gap, researchers use unsupervised domain adaptation, a method where models trained on well-lit data are adapted to handle low-light settings without needing manual body-joint labels. In a 2026 study, researchers generated realistic low-light training images to improve performance on low-light benchmarks.

Human Behavior and Resources

Another major hurdle is the sheer variety of human actions, especially when people interact with everyday objects like chairs or cups. The number of objects and the ways humans use them are too large to collect and label data for every possible combination. Researchers address this by exploring how foundation models and large vision-language models help systems recognise previously unseen interactions without prior labeled training. Beyond these behavioral challenges, today's most powerful AI models depend on massive computational resources in data centers with huge numbers of specialised computer chips. According to the 2026 AI Index Report, the technology industry produced more than 90% of the notable AI models introduced in 2025. These data centers are often too expensive for university labs or public-sector researchers, creating a mismatch between what is possible in theory and what smaller hardware can run in practical deployment.

Context

Computer vision and AI tools are increasingly being deployed in everyday environments like healthcare and transport to support human safety and well-being. However, practical hurdles such as unpredictable lighting, varied human actions, and high computing costs mean researchers must focus heavily on making these models robust rather than just accurate in ideal conditions.