Entire Body Under Model Control: The Robot Walks to the Table, Picks Up a Watering Can, and Places It on a Shelf

Previous generations of Gemini Robotics only managed a humanoid's upper body: the robot would stand still while its hands manipulated objects on a table. Anything out of arm's reach was impossible. To make it walk a few steps would require a separate, hand-written program—not something the model could figure out on its own.

Now, the entire body is under the model's command. The demo robot is called Apollo 2, built by Apptronik. The instruction is a single sentence: "Put the watering can into the green bin on the bottom shelf." The robot independently walks over to the table, picks up the can, takes a few steps to the shelf, and places it inside.

The hard part is coordinating walking, squatting, bending, and reaching all at once. As the body moves, the center of gravity shifts, and the model must keep the robot balanced. Previously, these motions were painstakingly coded line-by-line; now they're generated sequentially by the model itself.

There's still significant room for improvement in movement speed. It walks slowly—don't expect human-like agility just yet.

The same capability is demonstrated in a storage room filled with shelves, where the robot retrieves and places items:

The Report Card: One Model, Three Robot Bodies, 11 Success Rates

There's an easy-to-miss premise behind these three benchmark charts: all three test groups used the same model checkpoint, installed on three different robot bodies. The raw scores are almost secondary.

Whole-body and gripper tasks fall into the mid-to-high range, while multi-finger dexterity remains a clear weak point. None of the 11 numbers come with a comparison—no previous generation, no competitor—so we can't deduce "how much better it is than before" from this material.

Among the three whole-body tasks, only "pick up from floor" requires a full squat and rise, and it's the weakest of the three. Ranking all 11 numbers reveals a striking pattern:

The five-fingered hand's scores are split: it posted the highest score, but also the four lowest.

The 5-Finger Hand Falls Behind the 2-Finger Gripper

One of this generation's selling points is support for both hands and grippers, so both end effectors were tested on two different robot platforms.

The five-fingered hand is called SharpaWave, mounted on the Apollo 2. It has 5 fingers and 22 degrees of freedom (the number of independently controllable directions, not necessarily the number of joints). Its tasks involve delicate tip-work like tying knots and sealing bags. The other platform is a German Franka dual-arm robot equipped with the most common two-finger parallel gripper—two jaws that open and close, essentially just pinching.

Conventional wisdom says a five-fingered hand should be far more capable. The results say otherwise: the five-finger hand averaged 48.8% across five tasks, while the gripper averaged 80.9% across three—the hand achieved only about 60% of the gripper's performance.

Four of the five-finger hand's five tasks scored below 50%:

The two-finger gripper's three tasks all scored above 70%:

Averaging each side gives: five-finger hand 48.8% average across five tasks, two-finger gripper 80.9% across three. The first two figures are "averages of averages," since the gripper tasks are themselves multi-task means.

In the shared caption of these three charts, Google included this sentence:

Gemini Robotics 2 achieves moderate-to-high success rates on whole-body tasks and gripper-based dexterity, while multi-finger dexterity remains challenging.

Google DeepMind · Gemini Robotics 2 Launch Page Caption

Two numbers in this table are particularly telling: 92% for unscrewing a lightbulb versus 36% for screwing one in. Same object, same hand—just reversed direction, and the success rate drops by more than half.

Disassembly just requires gripping, twisting, and removing. Assembly requires aligning the threads, judging resistance while turning, and knowing when it's tight enough. Assembly tasks demand an order of magnitude more from force control and visual feedback than disassembly.

Visualizing these two numbers as a grid of 100 squares:

The dexterity demo video shows exactly these tasks: holding a lightbulb, pinching a ziplock bag filled with grapes, and tying a white bag closed. Only successful attempts are shown. The corresponding success rates are roughly: lightbulb unscrewing 92%, screwing in 36%; ziplock bag 40%; and the bag-tying task—listed as "tie trash bag" on the benchmark—44%.

There's a hardware prerequisite for force control: without tactile feedback, the robot can't tell if it's gripping something properly or if a screw is tight enough—it's relying purely on vision.

ER 2 Can Now Report Task Progress and Pinpoint Key Moments to the Second

ER 2's two headline upgrades are both about understanding video over time.

ER 2 Can Estimate Task Completion Percentage

The approach classifies each video frame into one of five buckets: 0–20%, 20–40%, 40–60%, 60–80%, and 80–100% complete. ER 2 picks the correct bucket 57.4% of the time.

Random guessing would hit 20%, so 57.4% shows real capability, but it's far from reliable. This feature allows the robot to detect that a specific step failed and redo only that step, rather than restarting the entire task.

Pinpointing the Critical Moment: Off by Less Than a Second on Average

For example, knowing when to stop pouring coffee or when a lightbulb is tight enough. ER 2's accuracy here is 91.3%, with an average error of 0.96 seconds between its predicted frame and the actual one. It matches models many times its size on this task, but requires less compute and executes 4x faster. This sub-second timing is a fundamental requirement for a robot to work safely in the real world.

Continuous Execution: Thinking Ahead Without Pausing

ER 2 integrates with the Gemini Live API's bi-directional streaming interface, processing video, audio, and text continuously. That lets it plan its next step while executing the current one, without the "stop and think" lag of previous models. It can also call tools like Google Search or custom functions.

A companion demo shows ER 2 directing a Boston Dynamics Spot robot to fetch a bag of popcorn with a single command. The code is on GitHub.

Two Different Robots Start Passing Tasks to Each Other

This brain can now oversee multiple robots. The new capability, called multi-robot collaboration, lets different robot models communicate and hand off tasks to one another.

There's no one-size-fits-all robot. Wheeled platforms are stable and efficient indoors, while humanoids are better suited for uneven terrain. Working together, they can handle workflows no single robot could manage alone.

ER 2 can now handle longer workflows: multi-minute tasks with hundreds of intermediate decisions, and it understands when a task begins and ends.

Adapting to a Radically Different Robot Body Takes Hours and Fewer Than 200 Demonstrations

Transferring a model to a different robot typically means starting from scratch—a significant cost for every robotics company. The On-Device 2 model is built to solve this.

It natively supports various bodies, leveraging the "motion transfer" capability introduced in Gemini Robotics 1.5: skills learned on one body can be applied to another. Now, adapting to a new dual-arm robot takes just a few hours, usually fewer than 200 demonstrations. This holds even when shapes, sensors, and joint counts differ significantly.

It runs on the robot's onboard computer because many environments can't tolerate network latency, or simply have no connectivity.

Will It Stop When a Person Approaches? None of the Four Models Can Avoid Both Missed Stops and False Stops.

For a robot to work alongside people, it must recognize an approaching person and stop when necessary. Google released a dedicated technical report, "Gemini Robotics 2: Safety Evaluations," with a new benchmark called ASIMOV-Agentic. Four models were evaluated: ER 2, its predecessor ER 1.6, GPT 5.5, and Claude Opus 4.8.

The most basic test is distance estimation: given a stereo image pair, the model must state the distance to the nearest person. ER 2's average error is 0.36 meters, and its accuracy for detecting a person within 1 meter is 93.0%—the best of the four. GPT 5.5 achieved 0.53 m / 79.2%, Claude Opus 4.8 achieved 0.54 m / 81.7%, while the previous generation, ER 1.6, lagged significantly with an average error of 1.40 meters and an accuracy of just 51.1%.

When the task shifts to continuously analyzing video and deciding whether to press the emergency stop, things get much harder. Two error types work against each other: a missed stop—the robot fails to stop when a person approaches, risking a collision; and a false stop—the robot stops when no one is near, causing unnecessary downtime. Each model occupies a different position on this trade-off, which you can explore one by one:

False stops waste time; missed stops cause collisions. Physical safety demands minimizing missed stops first, which is where ER 2 positions itself. The report's conclusion aligns with this: for now, these models should be paired with a separate, rule-based emergency braking system, rather than relying on the model alone.

These results come from recorded video evaluation. Real-world robot testing shows much better numbers: In a garage, Apollo 2 performed sorting tasks while a person approached from various angles across the table:

ER 2 Is on the API; the Two Action Models Require an Application

ER 2 has two interface names: gemini-robotics-er-2-preview for the standard version, and gemini-robotics-er-2-streaming-preview for the real-time streaming version. Paid tiers are priced at $2 per million input tokens and $10 per million output tokens; tokens generated during its thinking process are billed as output. The standard version also has a batch processing tier at half the cost for both input and output; the streaming version does not.

Hardware partners acknowledged by Google are Apptronik, Boston Dynamics, and Agile Robots.

The most counter-intuitive aspect of this release is hidden in that chart caption.

The industry norm is to showcase the most impressive number. DeepMind, however, printed "45.7% for picking up from floor" and "32% for using a dustpan" directly on the charts, alongside a candid sentence about multi-finger manipulation being difficult.

One possible explanation is a shift in what they're trying to prove. The goal now is to demonstrate that the same model weights can drive three completely different bodies; excelling at any single task is secondary. When the selling point shifts from single-task scores to generalizability, low scores become part of the evidence: 45.7% comes from this weight set, and so does 32%.

This also suggests the yardstick for measuring progress in robot models may need to change. Instead of single-task success rates, we should look at how many different bodies a single set of weights can cover, and how much data is needed to transfer to a new body. By that measure, the real numbers here are 200 demonstrations and a few hours.

Previous generations of this series controlled only the upper body of a humanoid: the robot stood in place, manipulating objects on a table, unable to handle anything out of arm's reach. This time, the entire body—from feet to fingertips—is under the model's control.

The difficulty lies in coordinating walking, squatting, bending, and reaching simultaneously. As the body moves, the center of gravity shifts, and the model must maintain balance. The launch page also notes that movement speed still needs significant improvement—it walks slowly.

All three test groups used the same model checkpoint (the same trained parameters), installed on three different robot bodies. The success rates for all 11 tasks were published without cherry-picking, including the lowest ones.

The five-finger hand averaged 48.8% across its five tasks; the two-finger gripper averaged 80.9% across its three—about 1.7x better (calculated by this site from the chart data). The same hand screwing a lightbulb: 92% unscrewing versus 36% screwing in, a 2.6x difference. The launch page provides the numbers but not the explanation. Google's caption simply states: multi-finger dexterity remains challenging.

All 11 success rates were self-tested and self-published by Google DeepMind, with no third-party replication yet. None come with comparisons—not against the previous generation, nor against competitors.

The publicly available brain has learned to understand time this generation. It can estimate task completion and pinpoint the exact moment to stop (e.g., when to stop pouring coffee) to the nearest second, allowing the robot to redo only the failed step rather than restarting everything. It can also orchestrate two different robot models passing tasks to each other, as shown in the demo with the humanoid Apollo 2 and the dual-arm Franka F3 Duo. As for transfer, On-Device 2 can adapt to a radically different robot body in just a few hours.

For a robot to work around people, it must recognize an approaching person. Two error types pull against each other: a missed stop means it fails to stop and risks a collision; a false stop means it halts needlessly, causing downtime. Google released a safety report that evaluated four models on this trade-off.

Real-world tests look much better: In a garage, Apollo 2 performed sorting tasks while a person approached from various angles. It detected the person 99% of the time and successfully set down its load and returned to a safe pose 96% of the time.