Google Unveils Gemini Robotics 2 for Full-Body Robot Control

A Robot That Used to Sit and Reach Now Stands Up

Previous generations of the model focused on controlling a humanoid's upper body to handle tabletop tasks. This new model has a single model move the legs, torso, arms, and multiple fingers together. The subtitle the company chose for its announcement was "from feet to fingertips."

It's the equivalent of someone who used to sit in a chair tidying up only within arm's reach now standing up, crossing the room, and crouching down to pick up items from the floor. Walking, lowering the body, moving through space, and avoiding obstacles are now handled at the same level as picking up an object. 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 capabilities were added with version 1.5 in September, and ER 1.6 strengthened physical-world reasoning this past April. Now, legs have been added.

Three Models, With Brain and Body Split Apart

This announcement isn't about a single model — it's three. Each plays a different role.

Gemini Robotics 2 is a vision-language-action (VLA) model. It converts what the camera sees and what a person says 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 can take minutes to complete. Google DeepMind explained that if a step fails partway through a long task requiring hundreds of decisions, this model can backtrack and fix it on its own.

The relationship resembles navigation software and a driver. ER 2 decides "turn left next," and the VLA actually turns the wheel. According to AI Times, ER 2 is integrated with the bidirectional streaming-based Gemini Live API, reducing the "stop and think" latency seen in previous systems.

The third is Gemini Robotics On-Device 2, a lightweight model that runs inside the robot without a cloud connection. Google DeepMind said adapting it to a new bimanual robot takes just a few hours, typically requiring fewer than 200 examples. It's comparable to how switching to a different bicycle only takes a few hours to get used to.

Two Robots Working as a Team

Another capability unveiled alongside this is the ability for multiple robots to work together. Google DeepMind said it added a function to ER 2 that lets robots operate as a team. Instead of having a single robot clean up a messy room from start to finish, the task can be split with another robot to finish faster.

It's similar to a kitchen scene where one person preps ingredients while another turns on the stove. Setting the order and coordinating so the robots don't collide with each other falls under the planning that ER 2 handles.

Google DeepMind Published Its Own Success-Rate Table

What stands out is that Google DeepMind released success-rate charts alongside the announcement. The figures were measured by attaching the same model checkpoint to three different robot form factors, unchanged.

In whole-body manipulation (Apptronik Apollo 2 + Inspire Hand), the success rates were 68.4% for tabletop pickup, 76.3% for shelf pickup, and 45.7% for floor pickup — meaning tasks requiring the robot to crouch down are the hardest.

Five-finger manipulation (Apollo 2 + Sharpa Wave Hand) showed wide variance. Unscrewing a light bulb scored 92%, but screwing it in only reached 36%. Tying a trash bag came in at 44%, sealing a ziplock bag at 40%, and using 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 wrote directly in its chart description that "precise manipulation with five fingers remains difficult."

The model's ability to read task progress was also quantified. According to AI Times, ER 2 classifies video progress into five stages with a classification accuracy of 57.4%, while "moment finding" — pinpointing when a specific event occurred — scored 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, Google DeepMind 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 can predict if a task is even feasible in the first place, and whether it asks a human to step in when it's uncertain.

Google DeepMind said ER 2 is the safest of 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 bring the robot to a safe stop.

ER 2 is available to try directly in Google AI Studio and is offered in private preview on the Gemini Enterprise Agents platform. VLA and On-Device 2 are open only to early-access partners. AI Times reported the program targets more than 100 testing companies. The hardware partners named in the announcement are Apptronik, Boston Dynamics, and Agile Robots.

「Editor's Take」

First, a disclosure: we have not tested the Apollo 2 ourselves. What we've verified for this article is limited to Google DeepMind's announcement, its charts, demonstration videos, and domestic Korean coverage. Robot performance can't be judged from edited footage alone, so this analysis is based solely on the published figures.

Looking at those figures, what stands out most about this announcement isn't the performance itself but how that performance was disclosed. Robot announcements typically show only the polished, edited highlights. This time, however, numbers like 45.7% for floor pickup, 36% for screwing in a light bulb, and 32% for the dustpan task were included directly in the announcement. The company itself wrote the line "remains difficult."

The detail worth dwelling on is that unscrewing a light bulb scored 92% while screwing it in scored only 36%. To a human, these are nearly identical motions. But unscrewing just requires gripping and turning, while screwing in requires aligning the threads.

It's an alignment problem at a contact surface that isn't visible to the eye. The gap between a language model generating sentences and a robot producing results in the physical world is compressed into that 56-percentage-point difference.

There are two takeaways for domestic teams. First, what's actually accessible right now isn't the VLA but ER 2. Since it's open in Google AI Studio, teams can test video understanding and task planning in isolation without needing a physical robot. Teams with camera-equipped facilities or sites have room to experiment.

Second is the claim that On-Device 2 "adapts to a new robot body in a few hours with fewer than 200 examples." If this holds up in practice, it lightens what has been one of the heaviest burdens in robot deployment — having to collect data from scratch every time the robot changes. The typical sequence of choosing hardware first and then fitting software to it could flip, with teams instead choosing the intelligence layer first and the body second.

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 any robot can use" (as reported by AI Times). The phrase "intelligence layer" is the key. Google isn't trying to sell robots — it's trying to capture the layer that runs inside robots regardless of who built them. The 45.7% and 36% figures are both evidence that this layer is still thin, and a map of exactly where it needs to be thickened.