Level 1 — Absolute Beginner

Google has a new computer brain for robots. It is called Gemini Robotics 2.5. It helps a robot see, think and move.

Google tested it in two car factories. One factory is in Hungary. The other is in Germany. Both factories make Mercedes-Benz cars.

The robot is called Apollo 2. It has two arms and two hands. Each hand has five fingers. The robot picks up metal car parts and puts them in trays.

The robot did the job well. It was right 96 times out of 100. Google says the robot works next to people. Google tested the brain on more than 200 different jobs.

robot

A machine that can move and do work by itself.

factory

A big building where people and machines make things.

arm

The long part of a body or robot used to reach and hold.

finger

One of the small parts of a hand used to hold things.

metal

A hard material such as iron or steel.

tray

A flat open box used to carry or hold things.

test

To try something to see if it works.

job

A piece of work that someone or something does.

Level 2 — Elementary

Google DeepMind has announced Gemini Robotics 2.5, which it calls its most capable vision-language-action model so far. A model of this kind takes what a camera sees, understands an instruction in ordinary words, and turns it into movement.

The company tested the model on more than 200 different tasks. It says success rates were 30 to 100 percent higher than earlier versions, depending on the task, and that the robots finished jobs 20 to 30 percent faster.

The most interesting test took place on real production lines. Working with the robot maker Apptronik and with Mercedes-Benz, Google ran a trial at plants in Kecskemet in Hungary and Sindelfingen in Germany. The humanoid robot Apollo 2 did a task called kitting, which means taking parts out of large bins and arranging them in trays in the order each car needs.

Google reported a 96 percent success rate for that task. Jeff Cardenas, co-founder and chief executive of Apptronik, said this is what Apollo was built for, meaning useful work beside people in real industrial settings. Readers should note that all of these numbers come from Google's own tests, and the company has not published how many attempts were measured.

model

A computer system trained on data to perform a task.

instruction

A statement telling someone or something what to do.

task

A piece of work to be completed.

success rate

The share of attempts that work correctly, shown as a percent.

production line

A line of machines and workers that build a product step by step.

humanoid

Shaped like a human being, with a head, arms and legs.

kitting

Collecting the right parts together in the order they will be used.

bin

A large open container for holding materials or parts.

Level 3 — Intermediate

Robotics demonstrations usually happen in laboratories, where the lighting is good and nothing is on a deadline. The more notable claim in Google DeepMind's release of Gemini Robotics 2.5 is that the model was put to work on active Mercedes-Benz production lines, at plants in Kecskemet, Hungary, and Sindelfingen, Germany, alongside human staff.

The model is described as a vision-language-action system, meaning it converts camera input and plain-language instructions directly into motor commands. Two models share the labour: Gemini Robotics ER keeps track of the overall workflow, while Gemini Robotics 2.5 controls the arms and hands. The hardware in the trial was Apptronik's Apollo 2 humanoid, fitted with a five-finger hand offering 22 degrees of freedom.

The task chosen was automotive kitting, which is less glamorous than it is revealing. Apollo 2 had to reach into bulk bins, identify metal brackets, and place them into sequencing containers in the order each vehicle on the line required, correcting for gaps as it went. Google reported a 96 percent success rate, along with broader gains of 30 to 100 percent over earlier baselines across more than 200 distinct tasks, and completion times 20 to 30 percent faster.

There are reasons for caution. Every figure is self-reported, no independent evaluation is described, and the company has not disclosed how many trials produced the 96 percent number. Google did publish safety results, claiming that reasoning guidance at the moment of action improved safe task completion by 33 percent around obstacles and 43 percent near people, which is the metric that will matter most if these machines are to share a factory floor.

demonstration

A public showing of how something works.

motor command

An instruction that makes a machine's joints or wheels move.

workflow

The sequence of steps needed to complete a job.

degrees of freedom

The number of independent ways a joint or hand can move.

bracket

A metal part used to hold or support something.

baseline

An earlier result used as a point of comparison.

self-reported

Published by the party being measured rather than an outside checker.

obstacle

Something that blocks the way and must be avoided.

Level 4 — Advanced

The interesting number in Google DeepMind's announcement of Gemini Robotics 2.5 is not 96 percent. It is the location. Humanoid robotics has produced a decade of videos shot in controlled environments, and the discipline's credibility problem has never been capability so much as the gap between a demonstration and a shift. This trial, conducted with Apptronik and Mercedes-Benz at plants in Kecskemet and Sindelfingen, took place on live production lines with human colleagues present and a schedule to meet.

Architecturally, the system splits cognition from control. Gemini Robotics ER maintains a representation of the workflow, deciding what should happen next, while Gemini Robotics 2.5 executes the manipulation, driving an Apollo 2 humanoid whose five-finger hand provides 22 degrees of freedom. The task, automotive kitting, requires pulling metal brackets from unstructured bulk bins and sequencing them into containers matched to each vehicle, compensating for gaps and misplacements along the way. It is dull, variable, and precisely the category of work that has defeated conventional industrial automation, which excels at repetition and fails at ambiguity.

The headline metrics are generous: a 96 percent success rate on the kitting pilot, gains of 30 to 100 percent over prior baselines across more than 200 tasks, and completion 20 to 30 percent faster. They are also entirely self-reported, unaccompanied by trial counts or methodology, and published by a firm with an obvious interest in the narrative. A 96 percent success rate is a different proposition on a factory floor than in a slide deck, since the residual 4 percent is where the recovery procedures, the stoppages and the liabilities live.

The safety figures deserve more attention than they will get. Google reports that applying reasoning guidance at inference time raised safe task completion by 33 percent around obstacles and 43 percent in the vicinity of people, and it has released an open benchmark for adversarial robustness on which its newest model is said to resist 99.5 percent of attempts to subvert it, against 82 percent for the previous generation. Those are the numbers that determine whether a humanoid is an industrial tool or an insurance problem, and they, too, await verification by someone other than the vendor.

credibility

The quality of being believable or trustworthy.

cognition

The mental process of understanding and deciding.

manipulation

In robotics, the physical handling and moving of objects.

unstructured

Not arranged in a fixed or predictable order.

ambiguity

A situation in which something can be understood in more than one way.

methodology

The set of methods and procedures used in a study.

residual

The part that remains after the main portion is accounted for.