The real test for AI begins when it leaves the screen

Recently, more than 2,000 humanoid robots ran races, played football and demonstrated a rapidly expanding range of physical capabilities at the World Humanoid Robot Games in Beijing. The event showcased remarkable advances, including a humanoid that reportedly surpassed Usain Bolt’s 100-metre world record time. Yet some of the most revealing demonstrations took place away from the sporting arena. Robots tackled tasks ranging from firefighting and housekeeping to retail assistance, as organisers recreated factories, hotels and model homes to test how these systems perform in real-world environments. The message from Beijing was unmistakable: humanoid robots are moving beyond impressive demonstrations and towards practical applications.

The Games reflect a broader shift in artificial intelligence. For most people, AI is still something they experience on a screen, generating text, analysing information or helping to write code. Increasingly, intelligent systems are being asked to act in the physical world. That is an exciting development. Advances in machine learning, computer vision, sensing and computing power are allowing robots and autonomous systems to perceive their surroundings and perform tasks that would have been extraordinarily difficult only a few years ago. As these systems become more capable, the questions engineers need to ask are beginning to change.

From capability to reliability

One of the misconceptions surrounding AI is that an accuracy score tells you everything you need to know about a system. In reality, the same level of accuracy can be exceptional in one application and unacceptable in another. If an AI assistant produces a poor draft of an email, a human can correct it. If a robot makes a mistake while sorting products in a warehouse, the consequences may be manageable. The calculation changes considerably when intelligent systems are involved in surgery, aviation or autonomous transport. As AI moves into the physical world, engineers are designing not simply for capability, but for consequence.

That shift is already changing the questions engineers ask. Unlike people, machines interpret the world through cameras, sensors and other streams of data. If those inputs are incomplete, misleading or unexpected, the system may respond in ways its designers never intended. Understanding those responses is essential if we want AI to operate safely and reliably outside controlled environments.

In our own work, for example, we have explored how visual information can alter the way an autonomous system interprets its environment, and how changing a drone’s perception of where the ground is located can cause it to respond incorrectly. For engineers, these are not simply failures to avoid, but problems to understand. AI has made extraordinary progress over the past few years, but larger models and greater computing power do not eliminate every engineering challenge. We need to understand why a system works, where it reaches its limits and how it behaves when conditions change. That understanding is what allows us to build systems that are both more capable and more reliable.

Sometimes we learn those lessons the hard way. During one experiment with a quadruped robot, it suddenly began kicking people on our research team. The manufacturer had omitted an emergency stop button, so the only way to halt it was to disconnect its power. The incident reinforced a principle every engineer understands: intelligent systems need robust safeguards as well as intelligence. Every AI-enabled machine should have a reliable way to stop, reset or return control to a human operator if something goes wrong.

A former Chief Technology Officer at Sikorsky, one of the world’s leading helicopter manufacturers, once made a point that has stayed with me. When we presented AI-based research, he told me that in mission-critical aviation applications, knowing a system works almost all of the time is simply not enough. Engineers need to understand what happens in the remaining cases, because that is where lives may depend on the outcome. As AI becomes embedded in more critical systems, designing for reliability under real-world conditions will become increasingly important.

Educating tomorrow’s engineers

Students today are understandably enthusiastic about AI. They want to learn machine learning, computer vision, large language models and the tools reshaping engineering. Universities need to respond to that demand. At the same time, we should not confuse learning to use AI with learning to understand it. Mathematics, physics, control theory and systems engineering remain the disciplines that allow engineers to explain why a system behaves in a particular way, diagnose failures and design more robust solutions.

The challenge for universities is not to choose between traditional engineering and AI, but to combine them. That matters particularly in the UAE, where AI is already being embedded across the education system and national investment continues to accelerate its adoption across government, industry and research. The opportunity is to create a generation that is not only comfortable using AI, but capable of building, testing and improving intelligent systems with the judgement their real-world deployment requires.

Students need experience with the latest advances in AI, but they also need the engineering foundations that allow them to question systems, validate their outputs and understand their limitations.

The robots competing in Beijing demonstrated how far intelligent machines have already come. Their long-term success, however, will depend on something less visible than faster running or greater lifting power. It will depend on how safely, predictably and reliably they perform in the complexity of the real world.

Preparing the next generation of engineers for that challenge may prove just as important as developing the technology itself. As AI becomes part of everyday life, the countries that lead will be those that educate people capable of understanding intelligent systems, testing them rigorously and continually making them better.

Professor Anthony Tzes is Distinguished Professor in Artificial Intelligence and Professor of Electrical Engineering at NYU Abu Dhabi