Building AI for the Physical World Is a Different Engineering Problem

Developing artificial intelligence for the physical world presents a unique set of engineering challenges. When moving AI development out of the browser and into physical environments like warehouses, factories, construction sites, vehicles, and other settings, the traditional assumptions about software architecture no longer hold true.

The network connection disappears, sensors generate noisy data, hardware can fail, devices have limited computational power, and the environment can change in ways that the training data didn't anticipate. Most importantly, the output of the AI system can directly impact the real world, which adds a layer of complexity to the engineering design process.

The core challenge isn't merely about placing an AI model next to an IoT device; it's about creating a system that reliably connects the physical world to software and transforms imperfect observations into useful decisions. The architecture of an industrial AIoT system goes beyond just the model, encompassing the physical environment, sensors or devices, connectivity, edge processing, data pipelines, AI or analytics, application logic, operational decisions, and physical actions.

Each stage of this chain introduces potential points of failure, such as sensor drift, connectivity issues, data overload, incomplete information, and the need for human decisions based on AI outputs. It's crucial to remember that AI is just one component within the overall system.

When approaching an AI project, it's easy to get caught up in choosing specific technologies like computer vision, large language models, edge models, or predictive maintenance systems. However, the more important question to start with is what physical problem you're trying to solve or improve. For instance, in a warehouse where equipment is frequently misplaced, the real issue might be ensuring you know the location and movement of important assets, their history, and whether they're following the expected workflow.

This focus on the physical problem can significantly alter the system architecture. Instead of focusing solely on AI, you might need to incorporate location sensors, connectivity infrastructure, an asset database, event processing, anomaly detection, and interfaces for operations teams. AI could play a role in this system, but it doesn't necessarily need to be the sole solution.

Physical systems often have messy inputs, which can make the AI model's job more difficult. For example, a vibration sensor measuring industrial equipment might produce noisy, missing, or outlier data, suffer from sensor drift, varying sampling rates, temporary connectivity problems, or changes due to normal operating conditions.

Without proper handling, feeding this raw data directly into an ML pipeline won't yield accurate results. Therefore, data quality becomes a critical aspect of AIoT development, requiring validation, timestamp synchronization, buffering, preprocessing, device identity, and data lineage considerations. In essence, the data engineering layer becomes as vital as the model itself.

Deciding between cloud processing and edge processing isn't a straightforward choice either. While cloud processing offers advantages like more computational power, centralized models, easier updates, and aggregated information, it might not always be practical due to rapid response requirements, unreliable networks, high bandwidth costs, sensitive data, or limited computational resources.

A possible solution is to use a hybrid approach, where an edge device handles local processing and filtering of data, while the cloud manages more extensive analytics, fleet management, and model updates. The ideal architecture depends on the specific needs of the system, such as response time, network reliability, bandwidth constraints, and data sensitivity.

Ultimately, the engineering standard for physical systems must prioritize reliability over making impressive demonstrations.