Google's Robot Model Scored 92% and 36% on the Same Light Bulb. Only One Number Traveled.

July 31, 2026. Google DeepMind released Gemini Robotics 2 on Wednesday, and one number carried the coverage: 92 percent, the model's success rate at unscrewing a light bulb. The same bar chart, in the same announcement, shows the same model screwing a light bulb back in 36 percent of the time. Both figures are Google's own. Both were published side by side. Only one of them traveled.

That gap is the most useful thing in the release for anyone who is not buying a humanoid robot, because it is the clearest public example this year of the difference between a benchmark a vendor publishes and the number a market repeats. If you are evaluating an AI tool for your business, that difference is the whole ballgame.

What Google actually published

1. Three models, not one. Gemini Robotics 2 is a vision-language-action model that converts vision and language into motor control and can drive a full humanoid. Gemini Robotics ER 2 is an embodied reasoning model that plans multi step tasks and coordinates. Gemini Robotics On-Device 2 is the efficient version that runs locally on the robot.
2. Whole body control, for the first time. Earlier models controlled the upper body for table top work. This one plans from torso to legs, so a robot can walk to a table, pick up a watering can, walk to a shelf and place it.
3. Multi robot collaboration. ER 2 can split a long task across different robot types, and it tracks progress from continuous video so that when a step fails it resumes from the last correct step instead of restarting.
4. Fast adaptation to new hardware. On-Device 2 adapts to a new bi-arm robot body in a few hours, typically on fewer than 200 examples, even when the shape, sensors and degrees of freedom differ substantially.
5. A new safety benchmark. ASIMOV-Agentic measures whether the reasoning model refuses unsafe tool calls from the action model, predicts whether a task is even possible, and asks a human for help when it is uncertain.
6. Availability today. ER 2 is live in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The action and on-device models are limited to early access partners. Hardware partners named are Apptronik, Boston Dynamics and Agile Robots.

The chart the headlines cropped

Google published three success rate charts. Read all three and the picture changes.

On multi finger dexterity, using a five fingered, 22 degree of freedom hand on the Apptronik Apollo 2, the results are: unscrew a bulb 92 percent, tie a trash bag 44 percent, seal a ziplock 40 percent, use a dustpan 32 percent, and screw a bulb in 36 percent. On general whole body manipulation the model picks an object up from a shelf 76.3 percent of the time, from a table 68.4 percent, and from the floor 45.7 percent. On two fingered gripper work it is stronger: precise insertion 89.6 percent, tool kitting 78.9 percent, general pick and place 74.2 percent.

DeepMind did not hide any of this. Its own caption under the chart says multi finger dexterous manipulation "remains challenging." The company reported the range honestly. The coverage reported the maximum. That is the pattern worth learning, because it repeats everywhere, and most of the time the vendor is less candid than Google was here.

Why a 90 percent step is not a 90 percent workflow

Here is the arithmetic that turns a benchmark into a business decision. If a single step succeeds 90 percent of the time and your process has ten steps that each must succeed, the end to end success rate is 0.9 to the tenth power, which is about 35 percent. At 95 percent per step it is about 60 percent. At 99 percent per step it is about 90 percent.

Nothing about that math is specific to robots. It applies to a lead qualification flow that reads a form, enriches a record, scores it, routes it, drafts a reply, books a slot and updates the CRM. Vendors quote you a per step number because per step numbers look good. You care about the compounded number, because that is the one your customer experiences. Ask for it, or compute it yourself from the step count.

This is also why the highest value automation work is rarely the flashiest. Cutting a workflow from ten steps to six raises reliability more than upgrading the model does, and it costs less. When we scope an AI automation build, the first pass is almost always removing steps rather than adding intelligence.

What it means for operators

You are not buying a humanoid. You are buying software agents that answer calls, qualify leads, update records and send replies, and every one of them ships with a demo that works. Three habits protect you.

First, insist on the distribution, not the headline. Ask the vendor for success rates across the full set of tasks they tested, not the best one. If they only have one number, that is the answer to a different question.

Second, test the mirror task. Unscrewing and screwing a bulb look like the same skill and differ by 56 points. In software the equivalent pairs are everywhere: an agent that books an appointment beautifully and cancels one badly, one that adds a CRM record cleanly and merges duplicates terribly, one that handles a happy path call and falls apart on a reschedule. Always test the reverse of the demo.

Third, read what the vendor admits. Google wrote its weakest result into the caption. Most vendors bury theirs in a footnote, a status page or a changelog. Whatever they concede in public is usually the floor, not the ceiling, of the problem.

Three questions that get you the whole chart

Before you sign anything, send these. They take a vendor five minutes to answer honestly and are revealing when they cannot.

1. What is the success rate on the task closest to mine, measured on data you did not train on, and how many attempts is that measured over?
2. What is the lowest score in the same test suite, and what task produced it?
3. What happens on failure: does the system retry, escalate to a human, or fail silently, and where do I see it?

That third question is where most deployments actually break. A tool with an 80 percent success rate and a clean escalation path beats a 92 percent tool that fails quietly, because the first one is a process and the second one is a liability.

The safety benchmark software teams should read

ASIMOV-Agentic was built for robots, but the thing it measures is a software problem. It tests whether the reasoning model will refuse an unsafe tool call from the action model, whether it can predict that a task is impossible before attempting it, and whether it asks a human when it is uncertain.

Swap the robot for a business agent and the questions are identical. Will your agent refuse an instruction that would email the wrong list, delete the wrong record or dial a number on a do not call list? Does it know when a task is outside its scope? Does it have a defined way to hand back to a person? Most agent stacks assembled from off the shelf tools answer no, no and no, and the failure only becomes visible when it costs something.

The fix is unglamorous and it is the same one we build for every client: a written register of what each agent is allowed to do, a named owner, a hard stop on the actions that cannot be undone, and a human in the loop on anything that touches a customer for the first time. If you want that designed properly rather than assembled, that is what our AI engineers and our AI automation agency team do.

Google published a model that can walk across a room and put a watering can on a shelf, and published the fact that it drops things off the floor more than half the time. That combination is what a mature disclosure looks like. Hold every other vendor you buy from to the same standard, and read past the first number.