Why AI's advances in robotics aren't going to change your life just yet

The article, written by Jamie Condliffe and published by MIT Technology Review in collaboration with the research foundation Aventine, dissects the gap between the runaway enthusiasm surrounding humanoid robots and what the technology is actually capable of doing today. The starting point is Optimus, Tesla's humanoid robot, which has appeared in videos dancing, handing out popcorn, taking out the trash, vacuuming or pressing buttons on a microwave, but also falling flat on its back while delivering water bottles or struggling to iron a shirt. Elon Musk has gone so far as to claim that Optimus will be "not only Tesla's biggest product, but probably the biggest product ever created," with "human and then superhuman" dexterity, capable of automating almost any human job for barely $20,000 per unit, and available to the public by the end of 2027, as he predicted at the World Economic Forum in Davos in January.

Musk is not alone in that optimism: Marc Andreessen has said robotics could become "the biggest industry in the history of the planet," Jensen Huang (Nvidia) claimed in January that humanoid robots would match human capability this very year, and Morgan Stanley projects nearly 1 billion robots "that look and act like humans" by 2050, in a market valued at more than $5 trillion. These forecasts rest on the idea that the same AI advances that drove tools such as ChatGPT or Claude will allow robots to imitate human movement just as chatbots imitate human language.

The article, however, sets against that narrative the skepticism of much of the research community. Yann LeCun, one of those regarded as the "godfathers of AI," said in Davos that "none" of the companies building humanoid robots "have any idea how to make those robots smart enough to be useful." Jonathan Hurst, co-founder and chief robot officer of Agility Robotics and a professor at Oregon State University, stresses a key distinction: "it's very easy to make a robot that looks like a person," but "it's dramatically harder to make a machine that moves or behaves dynamically or physically like a person." The piece insists that conflating humanoid form with "generalist" capability (learning and executing multiple tasks) is misleading.

To explain the real state of the art, the article turns to Google DeepMind's lab and its experimental platform ALOHA 2 ("A Low-cost Open-source Hardware System for Bimanual Teleoperation"), little more than a pair of arms, grippers and cameras, which serves as a testbed for the Gemini Robotics AI system. Controlled by that model, ALOHA 2 can, for example, pack a lunchbox by placing bread in a bag, grapes in a container and zipping everything shut: a modest task, but one that represents an objective leap over what was possible barely three years ago.

That progress is explained by the evolution of so-called robot policies, which determine how a machine perceives its environment, plans its movement and executes a task. They used to be hand-coded, line by line; now they are delegated to AI systems. First came vision-language models (VLMs), capable of identifying a suitable object in a scene (for example, a cloth to wipe up spilled coffee). Then came vision-language-action models (VLAs), which add movement commands learned through human teleoperation, allowing a robot to interpret an order such as "close the laptop" and carry it out, provided it has seen a similar task before. Gemini Robotics is precisely a VLA trained on many hours of human demonstrations, which allows it to perform relatively complex tasks such as picking up peas with kitchen tongs or doing origami, but with a clear limitation: outside its training set, it fails easily. As Edward Johns, professor of robotics at Imperial College London, sums it up, a model like that can today do only "a few things here and a few things there," far from a truly generalist policy.

The obvious route to broadening that generality —more training data— runs into a structural problem: unlike text for large language models, there is no ocean of high-quality physical demonstrations. The available solutions are all imperfect: hiring people to generate teleoperation data (expensive and slow), training on videos of people carrying out activities (lower-quality data) or deploying robots in the real world to gather experience (not very safe or reliable outside the lab). Pannag Sanketi, who led robotics work at Google DeepMind and now has his own project, advocates a "multi-front" strategy combining all those sources. But Hurst calls the idea that data alone is enough "a fundamentally flawed premise": seemingly simple tasks such as "making coffee" involve an explosion of variables (different kitchens, different machines, different cups, different ways of handling water, milk and coffee), and achieving generality through VLAs alone would require "complete data coverage of everything [a robot] could ever do," that is, an almost infinite training set. LeCun goes further and dismisses the approach: "the [AI] methods that have worked for language do not work for continuous, high-dimensional, noisy data," the kind typical of robotics.

The alternative gaining ground in research is so-called world models, trained not so much on text as on video, three-dimensional scans and sensor data, with the goal of building an internal representation of physical reality precise enough to predict how objects move, collide, fall or deform. The promise is twofold: simulations faithful to real physics would make development cheaper and faster by reducing the need for physical testing, and robots equipped with these models could reason about their environment and anticipate the consequences of an action before carrying it out, instead of simply reacting. Companies such as Nvidia and Google are working on it, and investment capital is flowing toward prominent startups: World Labs, co-founded by Stanford researcher Fei-Fei Li, raised $1 billion in February and was acquired by AMD at the end of September for $8.2 billion; AMI Labs, co-founded by LeCun himself after leaving his post as chief AI scientist at Meta, also raised $1 billion in March. Even so, Li herself described the field as "nascent," with "foundational approaches still being built," and in June she set out the outstanding challenges in a publication. For now, the article concludes, world models are a promising area of research, not an immediate path to generalist robotics.

As a sign of where this could go, the report recounts in detail the case of Physical Intelligence (PI, or π), a startup based in San Francisco's Mission District that is pursuing a "universal brain" capable, in theory, of turning any robot into a generalist. In 2024 PI released π0, a VLA trained on its own collection of 10,000 hours of human teleoperation demonstrations plus several open datasets. Version π0.5, from spring 2025, broadened the variety of data (including labeled images from the internet), gaining versatility; π0.6, in autumn 2025, added reinforcement learning. Each update improved the model's repertoire, from slowly folding clothes to putting objects away in new environments or folding boxes with a higher success rate.

The most striking leap came with π0.7, in April 2026, which incorporates a "lightweight" world model capable of generating images of the steps needed to carry out a task, giving the robot a kind of snapshot of what to do next. According to PI, the model shows the first signs of "compositional generalization": the ability to perform skills it has never seen by recombining others learned during training. In one test the robot was asked to "put a sweet potato in the air fryer," a task it had never encountered before; in the demonstration video, the machine hesitates, fails several attempts and finally manages a reasonable effort, though without fully completing the task. Sergey Levine, a professor at UC Berkeley and co-founder of PI, is enthusiastic: "it's the first time we've convincingly seen that kind of compositional generalization, where we can basically ask the model to do tasks for which we didn't specifically collect or train on data, and it makes a passable attempt." Curiously, when investigating why the model managed it, the team discovered that the training material contained relevant labeled snippets of teleoperation data, including two examples of a human operator pushing an air fryer basket inward; those few fragments may have been enough for π0.7 to come close to completing the sweet potato task. The article stresses that, given such minimal exposure to air fryers, it is still unclear how impressive this generalization ability really is, although it offers a sense of how far cutting-edge research can go today.

The piece closes by pointing to the distance between the lab's faltering steps and the dazzling fluency seen in many demonstrations and viral videos of robots dancing or courteously serving drinks: in many cases, it warns, humans are controlling the robot or have carefully scripted its actions. The very robot that appeared on stage alongside Nvidia CEO Jensen Huang in March 2025, apparently responding to his instructions and following him, was in fact remotely controlled by what its creators called "a puppeteer backstage." Fully autonomous motion planning —the robot knowing where to go, especially in new and chaotic environments such as a construction site or an unfamiliar house— remains, to a large extent, an unsolved challenge. An even greater and related challenge is getting robots to tackle broader, more ambiguous tasks that combine several subtasks and require deciding how and in what order to carry them out: the difference between a robot that puts a plate in the microwave and one that responds to the order "make dinner" by exploring the fridge, chopping ingredients and turning on the stove. Google DeepMind's own attempt at something similar —asking its robot to inspect a kitchen and put all the ingredients for a mushroom risotto into a basket— has so far ended in failure.

On top of that, a genuinely useful robot has to get it right practically every time. Marc Raibert, founder of Boston Dynamics, sums it up with a line the article highlights: with VLAs, "people get very excited when their result goes from 50% success to 70%," but "70% success is like it doesn't work, right?" The report concludes by recalling that the few humanoids being tested today in real-world settings —Agility Robotics has hundreds of robots deployed in trials at GXO Logistics, Amazon and Schaeffler facilities, according to the company itself— carry out extremely limited tasks in highly controlled environments, still far from any notion of a generalist robot.

Note on the source: the content received corresponds to virtually the entire article published by MIT Technology Review, but the text cuts off abruptly in the last sentence (after mentioning Agility Robotics' deployments), so it is not possible to know whether the article included a closing section or additional conclusions beyond the point where the download was interrupted.