Generated by Codex with GPT 5.6 Sol XHigh

Techmeme surfaced WIRED’s July 30 report, “Google’s Gemini Can Now Stomp Around as a Humanoid Robot,” about Google DeepMind’s attempt to move general-purpose AI out of chat windows and into machines with arms, hands, and legs. The eye-catching demonstrations show humanoids tidying shelves, tying bags, and handling lightbulbs. The deeper advance is architectural: Google has assembled perception, planning, motor control, and adaptation into a layered system that can operate several kinds of robots rather than one machine trained for one fixed routine.

The release is a substantial step toward more adaptable robots, but not “physical AGI” in any practical sense. The systems still rely on task-specific training, many reported success rates remain well below reliability, and the most capable motor-control models are limited to early-access partners. Google’s own safety evaluation also finds a persistent gap between understanding a rule in words and acting safely in a changing physical environment.

One brain, several bodies

Gemini Robotics 2 is a family of three connected models. Gemini Robotics ER 2 is the high-level embodied-reasoning model: it interprets human instructions and video, makes a multi-step plan, communicates with the operator, and delegates actions. Gemini Robotics 2 is the vision-language-action model that converts instructions and camera input into motor commands. Gemini Robotics On-Device 2 is a smaller action model designed to run locally when a robot cannot depend on a fast or continuous network connection.

That division resembles a planner supervising specialized executors. The reasoning model can track a task for several minutes and hundreds of decisions, notice that a step failed, revise the plan, and coordinate multiple robots. The action model handles the faster, lower-level work of walking, balancing, reaching, and manipulating objects. Keeping those roles separate also gives the planner a chance to reject an unsafe or infeasible action before passing it to the physical controller.

The important generalization is across robot bodies. Google reports using the same main model checkpoint with Apptronik’s Apollo 2 humanoid fitted with two different kinds of hands and with a Franka Duo system using conventional two-finger grippers. The on-device model can be adapted to a substantially different two-arm robot with less than 200 examples collected over a few hours. This does not eliminate retraining, but it makes a new robot body look more like a configuration target than a completely new research program.

WIRED notes the boundary clearly: Google trained the system with a mixture of human teleoperation, video demonstrations, and simulation, and robots still cannot perform a broad range of complex tasks without relevant training. The advance is faster transfer and a wider reusable base, not instant acquisition of arbitrary physical skills.

The averages matter more than the highlight reel

Google’s published results make the release more credible—and less magical—than the videos alone. On an Apollo 2 with standard robotic hands, average success rates for broad whole-body manipulation ranged from 45.7% when picking objects up from the floor to 76.3% when retrieving them from a shelf. A Franka Duo with grippers reached 74.2% on general pick-and-place tasks, 78.9% on tool kitting, and 89.6% on precise insertion.

Multi-fingered hands remain much harder. The Apollo 2 with a 22-degree-of-freedom SharpaWave hand unscrewed a lightbulb in 92% of trials, but screwed one in only 36% of the time. It tied a trash bag in 44%, sealed a ziplock bag in 40%, and completed a dustpan task in 32%. Those are meaningful research results because they involve contact-rich motions that are difficult to program. They are also failure rates that would be unacceptable for an unsupervised household or industrial product.

The comparison points to a practical lesson. General-purpose physical intelligence will not arrive as a single benchmark number. A robot can understand a command, form a reasonable plan, and still fail because an object slips, a hand lacks precision, the floor-level reach destabilizes the body, or the action lies outside its training distribution. Progress in reasoning does not make mechanics, control, sensing, and recovery disappear.

Access is similarly uneven. Gemini Robotics ER 2 is available through Google AI Studio and in a private enterprise preview, so developers can experiment with the reasoning layer. The vision-language-action and on-device models that directly control robots are available only through an early-access program. Most outside users therefore cannot yet reproduce the complete demonstrations on their own hardware.

Safety cannot be delegated to fluent reasoning

Google accompanied the launch with a robotics safety evaluation and the ASIMOV-Agentic benchmark. The proposed safety layer asks the reasoning model to refuse prohibited tasks, stop when a human gets too close or hardware fails, recognize when the action model lacks the needed skill, and request clarification when an instruction or scene is ambiguous.

The results reveal why this layer is necessary but insufficient. Frontier models exceeded 96% accuracy when safety constraints were presented as ordinary text classification. Performance varied more when the same constraints had to be grounded in spatial coordinates or executable tool calls. In other words, saying that a robot should not lift an open cup is easier than correctly locating the cup and preventing the controller from touching it.

Human-proximity tests show an even sharper tradeoff. From a single pair of stereo images, the tested models detected whether someone had crossed a one-meter boundary with roughly 79% to 93% accuracy. In the agentic test, systems tuned to avoid unnecessary stops could miss more than 40% of genuine hazards. Systems tuned to reduce missed hazards to roughly 10%–15% stopped unnecessarily 15%–25% of the time. None achieved both near-zero missed hazards and near-zero false alarms.

Google also reports stronger results in a controlled garage experiment: 99% human-detection accuracy and 96% reliability in moving an Apollo 2 into a safe pose. But the report explicitly says it does not evaluate the certified hardware, redundancy, real-time guarantees, and other functional-safety architecture required for compliant deployment. It also excludes privacy and fairness from the present benchmark.

That distinction should shape how the launch is read. An AI planner may add useful semantic safeguards—recognizing an ambiguous request, predicting that a skill is infeasible, or interpreting a fault message—but it should sit above deterministic protections such as emergency stops, force limits, and speed-and-separation monitoring. A probabilistic model cannot be the sole component preventing physical harm.

Physical AI has entered the integration phase

Gemini Robotics 2 is interesting less because one humanoid can tie one kind of bag than because formerly separate research problems are becoming one system. Multimodal reasoning can now supervise whole-body control, hand actions, local execution, adaptation to new hardware, and cooperation between robots. That creates a plausible platform layer for robotics, closer to an operating system shared across bodies than a script written for a single machine.

The remaining constraints are visible in the same release: uneven manipulation success, dependence on demonstration data, restricted access to the action models, and safety tradeoffs that still require conventional engineering. The bottleneck is shifting from whether a model can produce an impressive physical behavior to whether a complete robot can perform useful work repeatedly, recover gracefully, and remain safe around people.

That is a healthier standard than the “physical AGI” label. Gemini Robotics 2 shows that general AI techniques are starting to transfer into the physical world. It also shows that intelligence is only one layer of a dependable robot—and that the closer AI gets to people, the less impressive one successful demo should be compared with thousands of boring, predictable, safely handled repetitions.