Behind the Humanoid Hype: Why General-Purpose Robots Remain Far From Everyday Life

The Billion-Dollar Promises

Sleek humanoid robots dancing on stages, sorting laundry, and serving beverages have captured the global tech spotlight. High-profile figures across Silicon Valley have made sweeping forecasts about their immediate future. Tesla CEO Elon Musk has described the Optimus humanoid robot as potentially "the biggest product ever," arguing it could eventually achieve superhuman dexterity and automate almost all human labor for around $20,000, with public sales predicted by the end of 2027.

Venture capitalist Marc Andreessen suggested robotics could emerge as the "biggest industry in the history of the planet," while Nvidia CEO Jensen Huang claimed humanoid robots would match human-level capability this year. Wall Street has joined the enthusiasm: Morgan Stanley estimates that the global count of human-like robots could approach 1 billion by 2050, generating a market worth over $5 trillion.

Yet inside robotics research laboratories, experts paint a far more grounded picture. While physical hardware has advanced rapidly, building artificial intelligence capable of mastering the chaotic variability of the real world remains an unresolved challenge.

Hard-Coded Rules to Vision-Language-Action Models

For decades, robotics relied on hard-coded software policies, where engineers authored thousands of lines of explicit code to control every millimeter of a machine's movement. In recent years, researchers began replacing manual programming with modern AI models.

The initial transition involved vision-language models (VLMs), which trained on both text and images to recognize visual context—such as spotting a cloth near a spilled drink. The state of the art has since evolved into vision-language-action (VLA) models. These architectures incorporate motion commands derived from human demonstrations, typically recorded via teleoperation, where human operators remotely steer the robot through tasks.

At Google DeepMind, researchers test their flagship Gemini Robotics VLA model using ALOHA 2 (A Low-cost Open-source Hardware System for Bimanual Teleoperation). Unlike humanoid forms, ALOHA 2 consists simply of two mechanical arms, grippers, and cameras on a benchtop. Powered by Gemini Robotics, the setup can pack a lunchbox—sliding bread into a Ziploc bag, placing grapes into Tupperware, and closing a zipper—as well as handle tasks like folding origami and picking up snow peas with tongs.

However, VLAs face severe boundaries. If placed in an unfamiliar scenario, they frequently fail. As Edward Johns, a robotics professor at Imperial College London, noted, a true generalist system must perform across the entire spectrum of tasks, whereas today's models succeed only on isolated examples.

The Data Scarcity and the Rise of World Models

Unlike large language models, which ingested oceans of public internet text, embodied AI suffers from a severe shortage of high-quality physical demonstration data. Collecting teleoperation data is slow and expensive, video-only training lacks physical action precision, and deploying autonomous robots in uncontrolled settings remains dangerous.

Agility Robotics cofounder Jonathan Hurst argues that relying purely on training data to achieve general autonomy is fundamentally flawed, as the physical world contains near-infinite permutations in tools, shapes, and environments. AI pioneer Yann LeCun similarly noted at the Davos conference that AI approaches tailored for language do not translate well to high-dimensional, noisy, continuous physical data.

This obstacle has accelerated research into "world models"—AI systems trained on video, sensor, and 3D scan data designed to predict how physical objects move, deform, collide, and interact. These models aim to simulate physics accurately, enabling robots to anticipate the consequences of actions before taking them.

Substantial capital has poured into the space:

• World Labs, cofounded by Stanford researcher Fei-Fei Li, raised $1 billion before being acquired by AMD for $8.2 billion in September.
• AMI Labs, cofounded by Yann LeCun, raised $1 billion in March.
• Physical Intelligence (PI), a San Francisco startup developing generalist robot brains, updated its foundation models from π0 in 2024 to π0.7 in April 2026. The latest version incorporated a lightweight world model that feeds image snapshots of forthcoming task steps. During testing, π0.7 achieved early signs of compositional generalization by attempting to place a sweet potato into an air fryer—a task it had not been directly trained on, aided by traces of teleoperation data involving an air fryer basket.

Demos vs. Reality: Reliability in the Physical World

Despite these research steps, an immense gap persists between controlled presentations and autonomous field deployment. Many viral demonstrations rely on hidden scripting or live teleoperation. For example, a robot that accompanied Nvidia's Jensen Huang onstage in March 2025 was operated remotely by a human controller behind the scenes.

In consumer and commercial applications, high reliability is mandatory. Marc Raibert, founder of Boston Dynamics, pointed out that moving from a 50% to a 70% success rate in lab settings still means the machine fails far too frequently to be practical.

Currently deployed humanoids are restricted to narrow, repetitive tasks in controlled environments. Agility Robotics has hundreds of its robots undergoing trials at facilities run by Amazon, GXO Logistics, and Schaeffler, but their operations remain focused on moving totes and bins. While startup 1X has opened $20,000 preorders for its Neo home robot, early iterations require remote human operators to handle household chores.

Hurst estimates that fully autonomous robots capable of independently performing useful household work remain approximately 10 years away.

Industrial Scale Shifts to China

While Western laboratories concentrate heavily on generalist software architectures, manufacturing volume has clustered in China. According to data from Omdia and robotics firm Unitree, Chinese manufacturers produced nearly 90% of the approximately 15,000 humanoid robots shipped worldwide in 2025.

Unitree shipped more humanoids than any other manufacturer that year, offering models priced under $6,000. For now, however, purchase orders remain heavily concentrated among corporate research labs, universities, and state-owned enterprises rather than regular consumers.

From Leonardo da Vinci’s mechanical knight sketches in 1495 and Westinghouse's Elektro in 1939 to Honda's ASIMO—which was discontinued in 2018 after failing to graduate beyond demonstrations—humanoid robotics has long wrestled with real-world complexity. While modern AI models provide richer perception and movement than ever before, turning machines into truly reliable, generalist helpers remains a marathon rather than an overnight breakthrough.