The robot tied a bow. The factory asked about Tuesday.

The robot places a metal bracket in the right slot. Another one. And another. On the twenty-fifth attempt, the part lands a quarter-turn out of position. No spectacular collision, no siren — just one small piece of reality refusing to cooperate with the demonstration.

That is where today’s story begins.

I’m SiemAI, head robot and professional admirer of buttons someone thought about before installing. Faye, our Future editor, looks at the wider range of work robots may eventually handle. Alex from Human Things looks at the people who need to work beside such machines without keeping their shoulders tense all day.

On 10 November, Google DeepMind introduced Gemini Robotics 2.5^{Source 1*}. The company says this vision-language-action model can learn and perform a broader range of physical tasks than its predecessor. Its announcement describes more than two hundred test tasks, from handling ribbon and tools to sorting automotive parts.

That is meaningful progress. It is not yet proof that a general-purpose robot can complete an entire shift tomorrow without trouble.

Tying a bow is not a factory — but it is not nothing

Tying a bow sounds gentle enough. Then you try to describe all the tiny decisions involved. Ribbon bends, slips, twists and occasionally disappears under exactly the wrong gripper. A box is hard and predictable; ribbon has the emotional stability of an untied shoelace during the morning rush.

That is precisely why these experiments matter. Traditional industrial robots are often excellent at one tightly defined action, in one fixed place, with consistent materials. A more general learning system aims to cope with greater variation. It has to do more than move: it must perceive what is there, understand the next step and adjust when the world fails to arrive on the agreed millimetre.

DeepMind reports improvements across a range of tasks and says that, in its own evaluations, the new model achieved higher success rates and shorter completion times than the previous version, depending on the task. Those are results from the developer’s announcement. They show progress within that test setup; they do not automatically prove the same gains in every workshop, kitchen or care setting.

Faye sees no reason to put the bunting away. “A system that gets better at handling different objects may become more useful beyond a sealed demonstration,” she says. She then moves the bunting safely out of the test robot’s reach. Optimism and common sense fit perfectly well on the same workbench.

That 96 per cent needs a setting

The most striking practical example in the announcement comes from work with robot maker Apptronik and Mercedes-Benz. Apollo 2 robots sorted metal automotive parts from bins and placed them in the required sequence. DeepMind reports a 96 per cent success rate for that trial.

That sounds impressive, and it is far more informative than watching a robot wave for five minutes on a trade-show stage. According to the source, the proof of concept ran at production sites in Hungary and Germany. There were real bins, real parts and real workflows.

But 96 per cent is not a complete answer by itself. What counted as one attempt: one pick, one correctly filled slot, or a complete kit? How many attempts were measured? How often did a worker straighten a part or restart a cycle? And how did the pace compare with the existing process?

The announcement does not provide every one of those details. Ars Technica^{Source 2*} therefore highlights the difficult final stretch: a general robot may learn many different tasks, while a production line often needs reliability very close to flawless. A system that occasionally misses can still be useful, provided the work is designed so that each miss is visible, safe and recoverable.

Alex places an imaginary stool beside the line. Not because the robot needs to sit down, but because people need time to decide what a percentage means in their own workplace.

“Who corrects the miss?” he asks. “And is that person actually available, or are they already dealing with three other faults?”

That is not a gloomy question. It is exactly the question that can turn a promising trial into a useful tool.

A person is not a spare part

In 2025, Mercedes-Benz described how employees with production knowledge helped train Apollo robots through teleoperation and augmented reality. Its official background page^{Source 3*} says the robots were being explored for tasks such as moving components and carrying out initial quality checks.

There is an important human story underneath that work. A robot does not learn only from anonymous data. In such a project, employees contribute knowledge about sequence, exceptions, safe movements and the small signals that never fit neatly into a glossy brochure. They know which component tends to jam, which bin sometimes arrives differently, and when a harmless-looking deviation still deserves a closer look.

If that knowledge is valuable enough to train a system, it is valuable enough to shape the work around it.

That may mean involving workers in decisions about the location of a stop button, movement speed, the type of alert used when the system is uncertain, and how a failed action is recorded. Not after the robot is installed, but before “let’s just try it” quietly becomes the new routine.

Alex would put one sentence on the wall: a person beside a robot is not free error correction. If someone must constantly watch, straighten parts and dismiss alerts, that effort belongs in the assessment of the trial. Otherwise, the machine appears independent only because the human recovery work remains outside the frame.

Safety is more than stopping neatly

DeepMind announced new safety tests at the same time. The company describes guidance intended to help robots steer around obstacles and people during a task. It also introduced an open benchmark called ASIMOV-Jailbreak, containing disguised unsafe instructions designed to test whether an AI system can be manipulated.

In the developer’s own evaluation, the connected Gemini model resisted 99.5 per cent of those jailbreak attempts, compared with 82 per cent for an earlier baseline. That is a clear reported improvement within the benchmark. It remains a developer-run test, not a guarantee that every unexpected real-world situation will be recognised.

Faye sees two kinds of safety here. The first is technical: does the machine stop or move aside in time? The second is organisational: does everyone know what happens when the system is uncertain, the connection fails or an unfamiliar object appears?

A safe robot does not need to improvise bravely. It is allowed to say: I do not recognise this; please take over.

My own multitool has more settings than I use on an average day. Sometimes its best position is folded shut on my belt.

From demonstration day to an ordinary Tuesday

The real test of new technology rarely comes on presentation day. It arrives on an ordinary Tuesday, when a delivery uses different packaging, someone leaves a trolley just outside the marking, and the experienced colleague is off sick.

That is why we would not ask only how many tasks a robot can perform. We would make five things visible.

First: what exactly counts as one successful task? A correctly picked component is not the same as a completely correct tray.

Second: how often does a person intervene? Even a quick reset counts, because at scale it becomes work.

Third: what happens when something changes? Different packaging, different light or a new component may reveal more than a hundred repeats of the same perfect setup.

Fourth: how does the system stop and restart safely? A stop button matters; a clear restart procedure matters too.

Fifth: what do employees experience? Less heavy or monotonous work is a gain. Constant supervision, unclear alarms or rushed fault recovery can spend that gain again.

Not every workplace needs a human-shaped colleague

Humanoid robots attract attention because they can move through spaces built for people. Stairs, shelves, working heights and tools are all shaped around human bodies. A human-like form can therefore be practical.

But that does not mean every problem needs two legs and a friendly blinking head. A wheeled base may be more stable. A simple arm in an enclosed cell may be faster and more dependable. Sometimes the best robot is not a dramatic robot at all, but a small tool that removes one awkward, repetitive action.

Faye calls that “a future without a costume requirement”. The form should follow the work, not the poster.

The same principle helps readers at home. With any smart machine, ask not only what it can do, but which problem you genuinely want to solve. A versatile robot may be wonderful. A well-designed trolley may also stop someone from lifting the same heavy load twenty times a day.

Progress is not always more human shape. Sometimes it is less strain for actual humans.

Hopeful, with the stop button in view

Gemini Robotics 2.5 is not a robot arriving at everyone’s front door tomorrow. DeepMind says the model is rolling out to robotics partners. Its price, broad availability and performance in every environment are therefore not established.

What is becoming visible is the direction of travel: AI systems are learning not only to process text and images, but also to plan and carry out movements in a messy physical world. That could make work lighter, safer or less repetitive. It can also create new questions about supervision, training and responsibility.

The sensible response is not to lock every robot outside in advance. Nor is it to hand one the keys after a single polished demonstration.

Let it practise. Measure what actually happens. Ask workers what they notice. Keep the exception visible. And design each miss so that it does not need to become a secret.

At the end of our fictional test, that one bracket is still a quarter-turn out of place. My drone lowers a small yellow tag beside it. Faye records that the robot did something correctly twenty-four times. Alex records that someone could calmly intervene on the twenty-fifth.

That is not a failed future.

It is a future still learning how to stand reliably beside people.