AI breakthroughs in robotics won’t change your life any time soon

Humanoid robots resembling humans, such as Tesla’s Optimus, are increasingly visible in videos performing simple tasks like dancing, vacuuming, or handling objects. These robots are often presented as groundbreaking, with claims that they will soon automate most human labor at low costs. For example, Elon Musk, Tesla’s CEO, predicted in January 2026 that Optimus robots could be sold to the public by the end of 2027 and might eventually achieve superhuman dexterity. Other industry leaders, like Marc Andreessen of Andreessen Horowitz and Jensen Huang of Nvidia, have also expressed strong optimism about the future of humanoid robots. Morgan Stanley estimates that by 2050, nearly 1 billion humanoid robots could be in use, creating a market worth over $5 trillion. However, these robots currently struggle with basic tasks, often falling or failing to complete simple actions, highlighting the gap between hype and reality.

The optimism around humanoid robots is largely driven by advances in artificial intelligence (AI), particularly generative AI models like those behind tools such as ChatGPT. These AI systems are expected to enable robots to imitate human movement and adapt to new tasks. However, many robotics researchers, including Yann LeCun, a prominent AI expert, argue that current AI approaches are not sufficient for mastering the physical world. LeCun stated in January 2026 that companies building humanoid robots lack the knowledge to make them truly useful. The challenge lies in bridging the gap between AI trained on language and images and the complex, unpredictable nature of physical tasks, which require a different set of skills.

A key debate in robotics is whether humanoid robots—designed to look and move like humans—are necessary for general-purpose machines. Some researchers, like Jonathan Hurst of Agility Robotics, argue that creating a robot that moves like a person is far more difficult than making one that merely resembles a person. Generalist robots are designed to perform multiple tasks, but achieving this requires advanced AI systems. For instance, Google DeepMind’s ALOHA 2, a pair of robotic arms with cameras, demonstrates how AI can enable a robot to perform tasks like packing a lunchbox when controlled by their Gemini Robotics model. This system uses a vision-language-action model (VLA), which combines visual and language understanding with motion commands to execute tasks.

While AI models like VLAs have made progress, they have significant limitations. These models rely on training data consisting of videos or demonstrations of tasks, such as packing a lunchbox or folding laundry. However, they struggle when faced with tasks outside their training data. For example, a VLA-powered robot might fail if asked to perform an unfamiliar action. Edward Johns, a robotics professor at Imperial College London, notes that current models can only handle a limited range of tasks, describing them as doing "a few things here and a few things there" rather than achieving true generality. Researchers are exploring ways to expand this capability, but the lack of high-quality physical demonstration data remains a major hurdle.

Real-world tasks are far more complex than they appear. For instance, making coffee involves countless variables, such as different kitchen layouts, coffee machines, cups, and ingredients. Agility Robotics’ Jonathan Hurst argues that achieving true generality through VLAs would require complete data coverage of all possible tasks, which is nearly impossible. Yann LeCun dismisses the current AI approaches as inadequate for handling the high-dimensional, continuous, noisy data typical of robotics. He suggests that a different approach, such as world models, may be necessary. World models are AI systems trained on video, 3D scans, and sensor data to predict the outcomes of actions in the physical world, enabling robots to reason about their surroundings rather than just react to them.

World models are an emerging AI approach that aims to give robots a deeper understanding of the physical world. Companies like Nvidia and Google are investing in this technology, and startups such as World Labs and AMI Labs have raised significant funding—World Labs secured $1 billion in February 2026 and was acquired by AMD for $8.2 billion in September 2026, while AMI Labs raised $1 billion in March 2026. Fei-Fei Li, a Stanford AI researcher and cofounder of World Labs, describes the field as nascent and notes that foundational approaches are still being developed. Despite the challenges, world models represent a promising direction for achieving general-purpose robotics, offering the potential for faster, cheaper, and safer development by reducing the need for real-world testing.