Humanoid robots have moved from science fiction to real-world demonstrations, but they're revealing a surprising gap between athletic prowess and practical usefulness. At the World Robot Conference (WRC) 2026 in Beijing, Unitree Robotics showcased machines that can outrun Olympic sprinters yet struggle with everyday unpredictability. The event, which drew 300 exhibitors and 3,000 latest-generation robots, exposed both the promise and the profound limitations of embodied artificial intelligence (AI), the term for AI systems integrated into physical bodies.

Why Can Robots Break Speed Records But Not Do Laundry?

Unitree's newest model, called "Superman," reached a speed of 45.6 kilometers per hour on an athletics track during the concurrent World Robot Contest, surpassing Usain Bolt's 100-meter world record. However, the demonstrations revealed a critical flaw: the robots struggled to slow down after reaching maximum acceleration. Videos from the event show prototypes crashing into barriers and even short-circuiting, illustrating the gap between raw speed and controlled movement.

The G1 model from Unitree, by contrast, impressed observers with fluid movements and athletic abilities, suggesting that different design approaches yield different capabilities. Yet both machines highlight a fundamental challenge in robotics: speed and strength are easier to engineer than the nuanced decision-making required for household tasks.

What Are Robots Actually Doing Right Now?

The WRC 2026 showcased a diverse range of functional robots already deployed or ready for deployment. These machines demonstrate that robotics is moving beyond humanoid forms toward task-specific solutions. The variety of approaches reveals where robots are genuinely useful today:

• Warehouse and Retail Automation: The Galbot G1, an android on wheels with an elevating torso, autonomously replenishes shelves in supermarkets and pharmacies. A related model sorts 1,816 parcels per hour with 98% accuracy, making it effective in industrial and commercial settings.
• Heavy-Load Transport: Qiji, a robotic horse by Daka Robots, weighs 300 kilograms and can carry riders across rough terrain, demonstrating that non-humanoid designs excel at specific physical tasks.
• Domestic and Factory Work: X Square of Wall-B was designed for domestic environments to grab objects, collect waste, tidy up, and perform complex cleaning. With an additional metal arm, it transitions effectively into factory and industrial environments.

These examples show that robots are most reliable when performing repetitive, structured tasks under human supervision or in controlled environments. The challenge emerges when robots must adapt to unexpected situations.

Why Does Training Matter More Than Hardware?

A critical insight emerged from WRC 2026 forums: the bottleneck in robotics is not hardware design but training. Large language models like ChatGPT learn from vast amounts of text data available on the internet. Robots, by contrast, must learn physical perception through real-world experience. This difference means that scaling robot capabilities requires fundamentally different approaches than scaling AI models.

"It will take up to 10 years before we see robots capable of handling unexpected tasks with 80% autonomy," stated Wang Xingxing, founder of Unitree.

Wang Xingxing, Founder at Unitree Robotics

This timeline reflects the complexity of embodied AI. Robots must learn not just what to do but how their physical bodies interact with unpredictable environments. A robot that can sort packages in a warehouse operates in a controlled setting with known variables. A robot that must navigate a cluttered home, interpret human requests, and adapt to unexpected obstacles faces exponentially more complexity.

How to Understand the Current State of Home Robotics

The gap between laboratory demonstrations and household deployment reflects several practical realities:

• Controlled Environments Excel: Robots perform reliably on assembly lines, in warehouses, and in other settings where tasks are repetitive and variables are known. These environments allow robots to operate efficiently under human supervision.
• Unexpected Situations Cause Failures: Robots still appear uncomfortable within homes, where tasks vary, objects are arranged unpredictably, and human needs change moment to moment. The unpredictability that humans navigate effortlessly remains a major challenge for machines.
• Emotional Support Remains Experimental: UbTech's UWorld 1 humanoids, equipped with silicone skin and facial expression capabilities, demonstrated advanced interaction features including emotionally aware conversations and neck rotation during human engagement. Yet these remain prototypes rather than deployed solutions.

The WRC 2026 message was clear: humanoid robots are no longer promises or entertainment. They are functional machines solving real problems in specific contexts. However, the dream of a general-purpose household robot remains years away, constrained not by engineering ambition but by the fundamental challenge of teaching machines to learn from physical experience.

For investors and consumers watching Unitree and other robotics companies, the takeaway is nuanced. Robots that operate in structured environments with clear objectives are ready today. Robots that must adapt to human homes and unexpected situations require the kind of training timeline that Wang Xingxing described: a decade or more of real-world learning before they achieve the autonomy that would make them truly practical household companions.