A robot screws a bulb into a lamp, ties off a garbage bag, and drops a cassette into an old boombox. Watched cold, the clip is almost dull. That is the point of what Google DeepMind posted on July 30.
The company used the footage to introduce Gemini Robotics 2, the latest version of the AI it builds to run physical machines. The big change has an awkward name, intelligent whole-body control. Earlier Gemini Robotics software could already fold origami and seal a Ziploc bag, but it worked through robotic arms and hands operating on their own. This release lets a machine coordinate its whole body at once, walking, crouching, bending, and reaching while it works out what it is doing.
A brain that watches itself work
The launch centres on a model called Gemini Robotics ER 2, where ER stands for embodied reasoning. Google pitches it as a high-level brain. It does the thinking, plans a job that takes several steps, and decides what comes next, then passes the actual motor commands to a lower-level vision-language-action model that moves the limbs. ER 2 can also reach for ordinary tools by itself, Google Search among them, when a robot needs information it would not otherwise have.
What sets this apart from last year’s ER 1.6 is a sense of time. The model reads a running video feed of its own work and keeps track of how far along a task is, so it can catch a failure halfway through and retry a step instead of starting over. Google put numbers on this. On progress classification, which sorts each video frame by how complete the task is, ER 2 scores 57.4 percent. On moment-finding, catching the exact frame where something matters, such as the instant to stop pouring, it reaches 91.3 percent accuracy, and the company says it runs fast enough to keep the sub-second timing a working robot needs.
Robots that work in teams
The other new piece is teamwork. Google showed Apptronik’s Apollo 2 humanoid working next to a Franka FR3 Duo arm, the two machines splitting a task through a shared read of what needed doing rather than following one fixed script. A separate demo paired ER 2 with Boston Dynamics’ Spot, the four-legged robot, which went and fetched a bag of popcorn after a plain spoken request. The reasoning is that no single body suits every job. A wheeled robot is happy indoors, a humanoid manages rough ground, and a setup that lets them pass work back and forth can cover more than either one alone.
Developers do not have to wait to try it. Google has released ER 2 through the Gemini API and Google AI Studio, with a private preview running on its Gemini Enterprise Agent Platform. That last detail is the one investors will circle, because it moves the technology out of the demo reel and into the hands of outside engineers who can build products on it.
The honest asterisks
To its credit, Google was clear about the limits. The robots in the video are not all-purpose helpers. Each task was trained on its own, using a blend of human teleoperation, recorded video, and simulation, so the boombox trick does not carry over to loading a dishwasher. The company does insist the footage is real-time and fully autonomous, a jab you cannot miss at Tesla, whose flashy Optimus demonstrations were later shown to lean on humans operating the robots out of frame.
Safety gets heavier the moment a mistake involves a machine the size of a person rather than a wrong answer on a screen. Google says ER 2 is its safest robotics model so far, and that it will stop a humanoid when someone steps close and start again only once the area clears. It also published a new test it calls ASIMOV-Agentic, built to check whether a model can tell when a command would cause harm. “The safety question is even more pressing,” Carolina Parada, who runs robotics at Google DeepMind, told Wired, pointing to how many new situations these machines will land in.
Google keeps calling the goal physical AGI, a robot that can do anything a person can. That is still distant, and the company more or less admits it. Consumer robots are not shipping this year, and probably not next. What has changed is the shape of the thing. A system that can plan a job, watch its own hands, correct its own mistakes, and hand off to a second machine is a long way from a robotic arm folding paper on a bench. The question worth watching is which warehouse or factory floor gets to put it to work first.
Source: Google Just Unveiled Its Most Advanced AI Robots Yet – Gemini Robotics 2, TheAIGRID (YouTube).
For more coverage of robotics and AI, visit Mylistingo.
