이미지: AI 생성 — METAL LAB
"Pick up the watering can and put it in the green bin on the lower shelf."
Upon hearing this single sentence, the robot walked to the table. It picked up the watering can, took a few more steps to stand in front of the shelf, and set it down in the designated spot. This is a demonstration shown in the 'Gemini Robotics 2' announcement released by Google DeepMind on July 30 (local time).
As far as errands go, it looks trivial. Yet Google DeepMind wrote that this marked, "for the first time," full-body control of a humanoid robot.
A robot that once only sat and reached now stands up
Previous-generation models focused on controlling a humanoid's upper body to handle tabletop tasks. This new model has a single model jointly moving the legs, torso, arms, and multiple fingers together. The subtitle the company chose for its announcement was "from feet to fingertips."
It's as if someone who used to sit in a chair, tidying up only what was within arm's reach, has now stood up, crossed the room, and bent down to pick up objects from the floor. Walking, lowering the body, moving around, and avoiding obstacles are now handled at the same level as picking up objects. That said, Google DeepMind also added that "movement speed still needs further improvement."
Tracing the generations shows a clear direction. Gemini Robotics launched in March 2025, agentic concepts were added with 1.5 in September, and ER 1.6 strengthened physical-world reasoning this past April. And now, legs have been added.
Three models that split the brain from the body
This announcement isn't a single model but three, each with a different role.
Gemini Robotics 2 is a vision-language-action (VLA) model. It takes in what the camera sees and what a person says, and converts that into motor control signals. It's the one that actually moves the body.
Gemini Robotics ER 2 is an 'embodied reasoning' model that handles the robot's higher-level judgment. It converses with people, surveys a room, and plans multi-step tasks that take several minutes. Google DeepMind explained that this model can proceed through long tasks requiring hundreds of decisions, and if a step fails midway, it can go back on its own and fix it.
The relationship resembles that of a navigation system and a driver. ER 2 decides "turn left next," while the VLA actually turns the wheel. AI Times reported that ER 2 is integrated with the Gemini Live API, which is based on bidirectional streaming, reducing the "stop-and-think" latency seen in existing systems.
The third is Gemini Robotics On-Device 2, a lightweight model that runs inside the robot without connecting to the cloud. Google DeepMind stated that adapting it to a new bimanual robot takes just a few hours, typically requiring fewer than 200 examples. It's comparable to how it takes only a few hours to get used to riding a different bicycle.
Success rates disclosed by the company itself
What stands out is that Google DeepMind released success-rate charts alongside the announcement — figures measured by applying the same model checkpoint directly to three different robot forms.
For full-body manipulation (Apptronik Apollo 2 + Inspire Hands), picking up items from a table scored 68.4%, picking from a shelf scored 76.3%, and picking from the floor scored 45.7%. This suggests floor tasks that require lowering the body are the most difficult.
Five-finger manipulation (Apollo 2 + Sharpa Wave Hands) showed wide variance. Unscrewing a light bulb scored 92%, but screwing it in scored only 36%. Tying a trash bag followed at 44%, sealing a zip bag at 40%, and handling a dustpan at 32%.
By contrast, the two-finger gripper (Franka DUO) was relatively stable — 89.6% for precision insertion, 78.9% for tool kitting, and 74.2% for general pick-and-place. Google DeepMind itself noted in the chart description that "five-finger precision manipulation remains difficult."
The ability to read task progress was also given numerical figures. According to AI Times, ER 2 classifies task progress in video into five stages, from 0-20% up to 80-100%, with a classification accuracy of 57.4% and a "moment finding" accuracy—pinpointing the moment a specific event occurred—of 91.3%. This corresponds to the ability to know, while cooking, that "it's done now."
It stops when a person approaches
On the safety side, the company introduced a new benchmark called 'ASIMOV-Agentic.' It measures whether the reasoning model can refuse a dangerous tool call requested by the VLA, whether it predicts if a task is feasible in the first place, and whether it proactively asks a human to intervene when uncertain.
Google DeepMind stated that ER 2 is the safest among its robotics models on benchmarks for safety-constraint compliance and human proximity. If a person gets too close, it triggers a safety tool call to safely halt the robot.
What can be tried right now
ER 2 is available for immediate use in Google AI Studio and is offered as a private preview on the Gemini Enterprise Agent platform. VLA and On-Device 2 are open only to early-access partners. AI Times reported that more than 100 test companies are targeted.
The hardware partners named in the announcement are Apptronik, Boston Dynamics, and Agile Robots.
Editor's Note
Let's be clear about one thing first. We have not tested Apollo 2 ourselves. What we've confirmed for this article is limited to the announcement, charts, and demo videos released by Google DeepMind, along with domestic media coverage. Robot performance is not something that can be judged from edited video alone, so this piece is read purely against the disclosed figures.
Looking at those figures, what stands out most about this announcement isn't the performance itself but how the performance was disclosed. Robot announcements usually edit together only the best-performing clips. But this time, numbers like 45.7% for picking items off the floor, 36% for screwing in a light bulb, and 32% for handling a dustpan are included directly in the announcement. The sentence "still difficult" was also written by the company itself.
The detail that unscrewing a bulb scores 92% while screwing it in scores only 36% is worth dwelling on. To a human, the two actions are nearly identical. But unscrewing only requires gripping and turning, while screwing in requires aligning the threads.
It's a matter of fine alignment on an invisible contact surface. The gap between a language model generating a sentence and a robot producing a result in the physical world is compressed into that 56-percentage-point difference.
There appear to be two implications for domestic teams. First, what can actually be tested in practice right now isn't the VLA but ER 2. Since this model is open in Google AI Studio, teams can verify just the video-understanding and task-planning portions even without a physical robot. Teams with camera-equipped facilities or sites have room to experiment with it now.
Second is the claim that On-Device 2 "adapts to a new robot body in just a few hours with fewer than 200 examples." If this holds up in reproduction, it would lighten what has been the heaviest burden in robot deployment — having to collect data from scratch every time the robot changes. The order of first choosing hardware and then fitting software to it could be reversed, with the intelligence layer chosen first and the body chosen afterward.
Carolina Parada, Head of Robotics at Google DeepMind, said the goal is "to bring AI into the physical world and build an intelligence layer that every robot can use" (as reported by AI Times). The phrase "intelligence layer" is key. Google isn't trying to sell robots — it's trying to own the layer that runs inside robots, regardless of who makes them. The 45.7% and 36% figures are both evidence that this layer is still thin, and a map showing exactly where it needs to be made thicker.
