AI-rewritten: This is a summary of an article from MIT Technology Review, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

A collaboration between MIT Technology Review and Aventine reports on the rapid hype surrounding humanoid robots like Tesla’s Optimus. Elon Musk claims these machines will eventually possess superhuman dexterity and could be sold by 2027 for around $20,000 each. Other figures, including Marc Andreessen and Jensen Huang, have predicted that robotics could become the planet’s biggest industry or match human ability this year. Morgan Stanley estimates nearly one billion such robots could exist by 2050, creating a market valued over $5 trillion.

Despite this optimism, many researchers remain skeptical. Yann LeCun argues that current companies lack the know-how to make useful robots, while others note that existing AI struggles with the physical world’s infinite variability. There is also a distinction between robots that merely look human and those capable of performing complex tasks. Google DeepMind is currently testing advanced models like Gemini Robotics on simpler equipment called ALOHA 2, which uses two arms and cameras to perform tasks without needing a humanoid shape.

Progress in robotics relies heavily on "robot policies," which control how machines assess their surroundings and move. Historically, these were hard-coded rules written by engineers. Recently, advanced AI systems have taken over this role using vision-language-action models. These systems are trained on images, words, and motion data collected from humans performing tasks remotely. This allows the robots to learn new actions, such as packing a lunch or folding laundry, based on examples they have seen rather than pre-programmed instructions.

However, experts warn that current AI still has significant limitations. While models can perform complex tasks like origami or handling snow peas if trained on them, they are likely to fail if asked to do something outside their specific training set. The debate continues over whether these methods are sufficient for generalist robots or if a completely new path is required to achieve true physical intuition.