Tying a ribbon is barely worth noticing until you ask a robot to do it. Flexible material slips, fingers get in the way and one small mistake makes the whole knot collapse. On 10 November, Google DeepMind introduced Gemini Robotics 2.5, presenting demonstrations of delicate everyday tasks alongside an industrial experiment in sorting automotive parts. It is a vivid glimpse of what more adaptable machines might eventually manage.
The company's announcement describes a vision-language-action model, meaning a system designed to connect what a robot sees and understands with the movements it makes. Google also reports a 96% success rate in a car-parts kitting proof of concept with Apptronik's Apollo 2 robot in Mercedes-Benz production settings. That figure belongs to the tested task and setup; it is not evidence that the model can run a factory unattended.
Why a bow can be harder than a heavy box
For a human, lifting something heavy may demand strength, but tying a bow calls for constant tiny adjustments. The material bends unpredictably, tension changes while it moves and the next motion depends on the exact state of the knot. A machine cannot merely replay a series of arm positions and hope the cloth behaves as it did during training.
The same challenge appears in kitchens, workshops and care settings. Picking up cutlery, separating objects in a crowded drawer or positioning an awkward component requires awareness of the world as it changes. Robots have become effective at highly repeated tasks in controlled locations. What remains difficult is performing many different tasks in places that were not engineered around the robot.
Google's demonstration videos show the model attempting activities such as handling objects and bow tying. The eye-catching part is the variety, but a demonstration is not a full field trial. It shows particular successful behaviours under selected conditions. Buyers should want to know what happens when the ribbon is a different material, an object falls or a person interrupts the process.
The factory result is promising, with a precise boundary
The automotive test is more relevant to immediate commercial use. Kitting involves selecting and arranging parts so they are ready for an assembly task. It is not glamorous, yet mistakes can delay work or send the wrong component down a line. Google describes a proof of concept using Apptronik's Apollo 2 at Mercedes-Benz, with a reported success rate of 96% for the specified task.
That number is encouraging, but its meaning depends on the number of attempts, the difficulty of the examples and how failure was counted. A system that performs well on a known collection of objects could struggle when packaging changes, a component arrives damaged or a worker places an unexpected item in its path. A commercial deployment would need monitoring, recovery procedures and a clear division of responsibility.
The distinction between a research milestone and routine industrial operation matters because factory conditions are unforgiving. A human colleague must know whether the robot will stop safely, flag uncertainty and allow a straightforward override. A better grasp of objects does not automatically settle all the engineering and workplace questions around using the machine.
Gemini Robotics 2.5 must understand its own body
Earlier Gemini robotics work, explained in Google DeepMind's account of its previous-generation system, focused on combining physical movement with language understanding and whole-body control. The new generation aims to broaden the tasks a robot can learn and improve the speed with which it acquires them. In plain terms, Google wants machines to need less bespoke programming for each new activity.
That ambition resembles the shift from a special-purpose machine to a more flexible worker, although the comparison has limits. People bring common sense, awareness of consequences and years of physical experience to even a simple task. A model trained on pictures, instructions and demonstrations is still operating within particular data and hardware constraints. It may recognise a request yet misjudge how hard to grip an object.
There is a useful contrast with LiveAIWire's coverage of robot reasoning in simulation. Success in a virtual environment can help researchers learn faster, but reality adds friction, accidental contact and unusual lighting. Google is trying to bridge that gap by making perception and action work together in the physical world. The hard question is how reliably the model can recover when the world refuses to follow the script.
Learning a new task is not the same as mastering every task
A truly useful household robot would face a stream of unfamiliar situations: the washing is piled differently, the cupboard door sticks and the child has left a toy on the floor. The owner would not want to retrain the machine for each variation. That is why the ability to generalise from instructions and experience is valuable, even if the newest demonstrations remain far from that ideal.
Consider the research described in a separate LiveAIWire report on translating ordinary spoken instructions into robot actions. It illustrates a related difficulty: the words people use are often imprecise. Asking a robot to make the table look tidy leaves choices about what belongs where. A system can execute an instruction correctly and still misunderstand what the person actually wanted.
Those ambiguities become more serious around people. A machine handling cutlery must understand where hands are, what objects can break and when stopping is wiser than continuing. Reliability is not merely a percentage attached to the final completed task. It includes the manner in which the robot responds to uncertainty and prevents a small mistake from becoming a dangerous one.
The useful breakthrough would be an unremarkable Tuesday
Researchers understandably showcase the spectacular examples: a bow tied, a complex object grasped or a factory component handled without direct human control. But the customer who eventually pays for a robot will care more about a boring sequence of ordinary days. Does it complete the work, recover from mistakes and leave the workplace safer rather than more complicated?
Gemini Robotics 2.5 is a noteworthy step in Google's effort to produce adaptable physical AI. It should be judged as an advancing research and partner technology, not a retail product that solves every household chore. The real achievement will come when the clever demonstration stops being remarkable because the machine can be trusted to do the dull job well, again and again.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.