Generated by Codex with GPT 5.6 Sol High

Google DeepMind’s latest robots can change a lightbulb, fetch popcorn and tie a knot. Yet sweeping crumbs into a dustpan remains unreliable. That contrast captures the central problem in Adam Rogers’s article: artificial intelligence has become remarkably good at manipulating symbols, but a useful household or workplace robot must manipulate a physical world full of soft bags, fragile grapes, slippery glasses and unfamiliar rooms.

From programmed motions to physical AI

Traditional industrial robots work well because engineers tightly control both the machine and its surroundings. An arm on an assembly line can repeat a precise motion thousands of times, but the same approach becomes dangerous or useless in a home, where objects move and every task contains small surprises. DeepMind is pursuing “physical AI” instead: a general-purpose model that can transfer what it learns across jobs and even across robot bodies.

Gemini Robotics 2 extends this idea to whole-body movement. The software can control machines ranging from two-armed research platforms to Apptronik’s Apollo 2 humanoid. A higher-level model, Gemini Robotics ER 2, turns a spoken request into steps and estimates how far the task has progressed. A lower-level vision-language-action model converts those instructions and camera images into movements, revising its predictions four or five times a second.

The approach borrows from the history of AlphaGo and large language models. Reinforcement learning lets a robot discover a skill by trying repeatedly and receiving numerical rewards, but real-world practice is slow. Imitation learning is faster: people teleoperate a robot or record themselves performing a task, and simulations add still more examples. The hope is that enough varied data will reveal patterns that generalize, just as a language model learns relationships that were never explicitly programmed.

That hope has attracted enormous investment. A 2025 Morgan Stanley report projected a humanoid-robot market worth more than \$5 trillion by 2050; Tesla, BMW and numerous Chinese start-ups are pursuing their own machines. Yet the article places those forecasts beside a more telling reality: much of the industry’s future rests on robots that still struggle with ordinary chores.

The missing sense of touch

DeepMind’s demonstrations show both the promise and the gap. A small robot can reason through a request for popcorn, but the model classified task progress correctly only 57.4 percent of the time. Apollo 2 swept debris into a dustpan successfully in just 32 percent of tests. Bags and brush bristles deform, grapes bruise, and a glass must be gripped firmly enough not to slip but gently enough not to shatter.

Those failures expose a basic imbalance in the training data. The Internet contains vast quantities of images and text, but almost no records of force, touch or proprioception. A robotic hand may have 22 independent ways to move while receiving little information about what its fingers feel. Modern hardware can include torque sensors and strain gauges, but collecting enough tactile data to train a broadly capable model remains difficult. Visual intelligence alone does not supply the fast feedback humans receive through nerves, muscles, tendons and joints.

Safety therefore has to be layered rather than assumed. DeepMind’s model can stop a robot when someone gets too close, and the company has introduced an “Asimov” benchmark for dangerous behavior. Robot manufacturers add their own sensors and protective systems. These measures matter because a model can execute a locally reasonable motion while still misunderstanding the larger objective.

Why humanoids keep returning

A human shape is not necessarily the most efficient design for carrying loads or moving through rough terrain. It is attractive because the world is already built around human proportions: doorways, countertops and tools all favor bodies like ours. Human demonstrations are also easier for a similarly shaped machine to imitate. This creates a practical reason for humanoids that has little to do with science-fiction aesthetics.

The article closes with a philosophical reversal. Theories of embodied cognition argue that understanding an apple requires a body that can hold, smell and taste it. DeepMind’s work suggests that a digital model may learn a great deal about apples without those experiences. A robot might not need an inner life resembling a person’s to identify and handle one. But abstract knowledge is not enough to prevent crushed fruit or broken glass. Whatever intelligence emerges from more data, a dependable physical AI will still need something like nerve endings.