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

The pitch for humanoid robots has gotten very big. MIT Technology Review, in a piece produced with the nonprofit Aventine, lays out the claims: Elon Musk calling Tesla's Optimus possibly the biggest product ever and predicting public sales by the end of 2027, Marc Andreessen calling robotics potentially the biggest industry in history, and a Morgan Stanley projection of nearly 1 billion humanlike robots by 2050.

The reporting then walks through what labs can actually do. Vision-language-action models have replaced much hand-coded robot behavior, and systems like Google DeepMind's Gemini Robotics can pack a simple lunch or fold origami. But they tend to fail on tasks outside their training data. Researchers disagree sharply on whether more data will fix that or whether something new, such as world models, is needed.

This matters because investment, labor policy and consumer expectations are being set by the bold forecasts, not by the lab results.

We came away from this piece more optimistic about the science and more skeptical of the timelines, and we think that is the right combination. The progress is genuine. Handing robot control to learned models instead of thousands of lines of hand-written rules is a real shift, and the article's examples show robots doing delicate, multistep tasks that were out of reach a few years ago.

The most interesting moment is Physical Intelligence's π0.7, which the company says shows early compositional generalization: asked to load a sweet potato into an air fryer, a task it was not specifically trained on, the robot made a passable, unfinished attempt. Then the team found a couple of relevant air fryer clips buried in its training data. That is the whole debate in miniature. Was it reasoning, or remembering scraps? Nobody can yet say for sure.

The line that should anchor expectations comes from Boston Dynamics founder Marc Raibert, who points out that a jump from 50 to 70 percent success excites researchers, yet 70 percent success is effectively a robot that does not work. Homes and factories need reliability close to every time. Agility Robotics' Jonathan Hurst, whose company has hundreds of robots in trials moving bins and totes, guesses about 10 years before robots are doing useful things in people's homes.

The downside of the hype gap is concrete. Demos often hide human teleoperators, and the article notes the 1X Neo home robot, available for preorder at $20,000, needs a remote operator for most tasks, which means a person may be looking into a customer's home through its cameras. Consumers paying premium prices for partly puppeteered machines, and workers being told automation is imminent when it is not, are both real harms from overselling.

There is also a geopolitical thread. The piece cites Omdia and Unitree figures showing nearly 90 percent of the roughly 15,000 humanoids shipped in 2025 came from Chinese companies. Whoever wins on cheap, capable hardware may matter as much as whoever wins on the brain.

Our read: back the research, discount the launch dates. I'm watching for reliability numbers from real deployments, not demo videos, and for honest disclosure whenever a human is in the loop.

Humanoid robots have a long history of impressive demos that stalled short of usefulness, from Westinghouse's Elektro in 1939 to Honda's ASIMO, discontinued in 2018. The current wave differs in using AI models trained on images, language and robot motion data, though gathering enough high-quality physical training data remains a central obstacle.