Beyond the screen: Why AI's real challenge is reliability in the physical world

Beyond the screen: Why AI's real challenge is reliability in the physical world

At the second World Humanoid Robot Games in Beijing, 666 teams from around the globe competed with more than 2,056 humanoid robots across 51 events, representing a 138 per cent increase in participation from the inaugural 2025 competition. Among the highlights, the Tiangong Ultra robot, developed by Beijing Humanoid Robot Innovation Center, clocked 9.39 seconds in the 100-metre sprint, surpassing Usain Bolt's 9.58-second world record set at the 2009 World Athletics Championships in Berlin. The same robot had won the event just a year earlier with a time of 21.50 seconds, a striking illustration of how rapidly this technology is advancing.

Yet the most significant developments at the Games took place beyond the athletic arena. Organisers constructed replica environments including factories, hotels and residential spaces to test how robots handle firefighting, housekeeping, retail service and other practical tasks. Each event featured an Autonomy Weight Coefficient that determined the level of independent operation required, with many 2026 competitions demanding full autonomy across more than 1,000 individual contests. The conclusion was unmistakable: humanoid robots are moving from impressive spectacles to functional applications.

AI enters the physical realm

This shift reflects broader changes in artificial intelligence. Most people still encounter AI primarily through screens, where it generates text, analyses data or assists with programming. Increasingly, however, intelligent systems are being deployed in physical environments. Advances in machine learning, computer vision, sensing technology and computing power now enable robots and autonomous systems to perceive their surroundings and execute tasks that would have seemed prohibitively complex only recently. Real-world deployments already include logistics operations for material handling and package delivery, healthcare settings for patient assistance and hospital logistics, retail environments for customer service, and manufacturing facilities such as BMW's Spartanburg plant, which deployed Figure 03 humanoid robots in February 2026.

As these capabilities expand, the fundamental questions facing engineers are evolving. The challenge is no longer simply whether a system can perform a task, but whether it can do so safely and reliably under diverse conditions.

From performance metrics to consequence management

A common misconception about AI holds that accuracy percentages provide complete assessments of system quality. In practice, identical accuracy rates can be excellent in one context and inadequate in another. When an AI assistant produces a flawed email draft, human oversight can correct it. When a warehouse robot mishandles products, the consequences may be manageable. The equation changes substantially when intelligent systems participate in surgery, aviation or autonomous transportation. As AI assumes physical agency, engineers must design not merely for capability, but for the consequences of failure.

This evolution is already reshaping engineering priorities. Machines interpret their environment through cameras, sensors and data streams. When these inputs prove incomplete, misleading or unexpected, systems may respond in unintended ways. Understanding those responses becomes essential for safe operation outside controlled settings.

Our research has explored how visual information can alter an autonomous system's environmental interpretation, such as how changing a drone's perception of ground level can trigger incorrect responses. For engineers, these represent not simply errors to prevent, but problems requiring systematic understanding. Despite extraordinary recent progress, larger models and greater computing power do not eliminate every engineering challenge. We need to comprehend why systems function as they do, where their limitations lie, and how they behave when conditions change. That understanding enables building systems that are simultaneously more capable and more dependable.

Experience sometimes teaches these lessons harshly. During one experiment with a quadruped robot, the machine suddenly began kicking research team members. The manufacturer had omitted an emergency stop mechanism, leaving power disconnection as the only means to halt it. The incident reinforced a fundamental principle: intelligent systems require robust safeguards alongside intelligence. Every AI-enabled machine should incorporate reliable methods to stop, reset or return control to human operators when problems arise.

Engineering for mission-critical applications

A former Chief Technology Officer at Sikorsky, a leading helicopter manufacturer, once articulated a point that resonates strongly in this context. When presented with AI-based research, he emphasised that in mission-critical aviation applications, knowing a system works most of the time proves insufficient. Engineers must understand what occurs in remaining cases, because lives may depend on those outcomes. As AI becomes embedded in more critical systems, designing for reliability under real-world conditions will grow increasingly vital.

Preparing the next generation

Students today demonstrate understandable enthusiasm for AI, seeking expertise in machine learning, computer vision, large language models and the tools reshaping engineering practice. Universities must respond to this demand. Simultaneously, institutions should distinguish between learning to use AI and learning to understand it. Mathematics, physics, control theory and systems engineering remain the disciplines enabling engineers to explain system behaviour, diagnose failures and design more robust solutions.

The challenge for universities involves combining traditional engineering foundations with AI capabilities rather than choosing between them. This matters particularly in the United Arab Emirates, where AI integration spans the education system and national investment accelerates adoption across government, industry and research. The UAE government has committed over AED 1.5 billion to enhance digital education infrastructure through initiatives such as the Smart Learning programme, impacting more than 400,000 students. Institutions like NYU Abu Dhabi's Center for Artificial Intelligence and Robotics conduct fundamental research in areas including multi-agent systems, planning and navigation, deep learning for robotic safety and resiliency, and human-machine interfaces. The opportunity exists to cultivate a generation not merely comfortable using AI, but capable of building, testing and improving intelligent systems with the judgement their real-world deployment demands.

Students require experience with the latest AI advances, but they equally need engineering foundations enabling them to question systems, validate outputs and understand limitations.

The path forward

The robots competing in Beijing demonstrated how far intelligent machines have progressed. With the global humanoid robot market reaching approximately 18,000 units and USD 440 million in revenue in 2025, representing a 508 per cent year-over-year increase, and projections estimating growth from USD 5.41 billion in 2026 to USD 50.27 billion by 2035, the commercial trajectory is clear. Long-term success, however, will depend on factors less visible than enhanced speed or lifting capacity. It will depend on how safely, predictably and reliably these systems perform amid real-world complexity.

Preparing the next generation of engineers for this challenge may prove as important as developing the technology itself. As AI becomes woven into everyday life, leadership will belong to nations that educate people capable of understanding intelligent systems, testing them rigorously and continually improving them.

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