MIT Technology Review: robotics AI progress is real, but consumer humanoids remain far off

Hype meets hardware

A piece from MIT Technology Review, reported in collaboration with the non-profit research foundation Aventine, argues that recent AI breakthroughs in robotics — however real — will not reach consumers any time soon. The hype around humanoid machines, it says, is obscuring slower, painstaking progress in labs, and the boldest timelines from executives deserve heavy discounting.

The bullish case

The article opens with Tesla's Optimus, a white humanoid with a black head and torso that has circulated widely in video clips: dancing, vacuuming, taking out trash, pressing microwave buttons — and also toppling over while handing out water bottles. Elon Musk has told shareholders the robot will eventually have "human and then superhuman dexterity," and has called it not just Tesla's biggest product ever but probably the biggest product ever, according to MIT Technology Review. At Davos in January he predicted public sales by the end of 2027, at a price as low as $20,000 per unit.

He is not alone in the optimism. Marc Andreessen has said robotics could become the biggest industry in the planet's history. Nvidia CEO Jensen Huang said in January that humanoid robots would match human-level ability within the year. Morgan Stanley projects nearly 1 billion humanlike robots by 2050, in a market it values at more than $5 trillion.

Why researchers are skeptical

Much of this enthusiasm assumes the AI behind tools like ChatGPT and Claude can teach machines to imitate human movement the way chatbots imitate language. Many roboticists disagree, arguing that an intelligence built on language and images is a poor fit for the infinite variability of the physical world. Yann LeCun, described by MIT Technology Review as one of the godfathers of AI, said at Davos that none of the companies building humanoid robots — "absolutely none of them" — knows how to make them smart enough to be useful.

The article also flags a conceptual sleight of hand: conflating robots that look human with generalist machines that can learn and perform many tasks. Jonathan Hurst, cofounder and chief robot officer of Agility Robotics and a professor at Oregon State University, notes that building a robot shaped like a person is easy; building one that moves and behaves physically like a person is dramatically harder.

What is actually improving

Beneath the noise, the field has shifted in a fundamental way. Robot policies — the systems that determine how a machine assesses its surroundings, plans movement and executes a task — were once hard-coded by engineers, with thousands of lines of software dictating every millimeter of motion. Increasingly, those policies are handed to AI models instead.

First came vision-language models, which understand images as well as text and gave robots contextual awareness they lacked a few years ago. Now vision-language-action models, or VLAs, add motion commands, trained on task footage plus teleoperation data in which humans remotely guide a robot through actions.

MIT Technology Review points to Google DeepMind's Gemini Robotics, a VLA tested on ALOHA 2, a deliberately simple rig of two arms, grippers and cameras. Controlled by the model, it can pack a lunchbox: bread into a Ziploc bag, grapes into a container, both placed into the box and zipped up. It can pick up snow peas with kitchen tongs and fold origami. Modest as that sounds, the article calls it an objective step beyond what was possible even three years ago.

The generalization problem

The glaring limitation is that a VLA-controlled robot asked to perform a task outside its training set will very likely fail. Roboticists believe an equally transformative revolution to the one that produced generative AI is possible in robotics, endowing machines with physical intuition that has long been out of reach — but they disagree on when it will arrive, and on whether today's AI methods can get them there at all or whether an entirely new path is required.

Why it matters

Executive timelines shape investment, regulation and public expectations. If Musk's 2027 consumer date and Huang's near-term human-level claims drive decisions, the gap between promise and capability could trigger a backlash against a field that is, by the article's account, compounding real gains. For technical readers, the signal worth watching is not humanoid launch dates but the underlying architectural shift: from hand-coded policies to vision-language and vision-language-action models trained on human demonstrations. That is where the measurable progress is happening — and it suggests consumer impact will arrive gradually, through increasingly capable narrow tasks, rather than through a sudden humanoid leap.