The distinction matters because industrial robots are not judged like conversational AI systems. A model can appear impressive across hundreds of trials and still create unacceptable costs if it fails unpredictably during repetitive production. Google DeepMind’s Gemini Robotics 2.5 therefore arrives at a critical point for the robotics industry: laboratory progress is accelerating, but the final step toward reliability may be the most difficult one.

“You could manually program a robot to do a handful of tasks in a factory without issue-this type of automation has been common for decades. Gemini Robotics is not trying to just do a few things, though. These are intended to be general robots that can handle almost anything, but there are trade-offs to that approach.”

In its announcement, Google DeepMind said Gemini Robotics 2.5 was evaluated on more than 200 dexterity tasks, with success rates between 30% and 100% higher than previous baselines. The company also reported that the system completed tasks 20% to 30% faster than those baselines.

Those figures describe a broad improvement rather than one isolated capability. Dexterity tasks require a robot to perceive objects, plan movements and adjust when the physical world does not behave exactly as expected. That combination is central to general-purpose robotics. Factory environments are structured, but they are not perfectly predictable. Parts shift, packaging changes, lighting varies and objects can become misaligned.

The announcement’s most commercially significant result came from a proof of concept involving Apptronik’s Apollo 2 humanoid robots on Mercedes-Benz production lines. DeepMind reported a 96% success rate for automotive kitting, a process in which components are collected and prepared for assembly.

A 96% rate is notable because it shows the model operating in a real production setting rather than only in a laboratory. It is also revealing because the remaining 4% represents the central deployment problem. In a factory, a failure is not simply a missed benchmark attempt. It may interrupt a workflow, require human intervention or force a worker to monitor the robot closely enough to reduce the expected savings.

The gap between demonstration and dependable operation was the focus of Ars Technica’s Ryan Whitwam, who reported critically on the release. Whitwam wrote that the factory result is likely below the reliability expected for repetitive production work. He also quoted DeepMind’s Kanishka Rao describing the final 4% as a “steep road.”

That criticism does not erase the significance of the result. It clarifies what the result means. Gemini Robotics 2.5 may be strong enough to assist with tasks where humans can correct occasional mistakes, but that is different from replacing a tightly controlled industrial process. The commercial value of a robot depends not only on its peak performance, but also on the frequency, cost and consequences of its failures.

The model’s architecture raises another question. Ars Technica reported that Gemini Robotics 2.5 runs in the cloud, which introduces potential latency concerns. A delay may be manageable when a robot is sorting parts at a relatively slow pace. It becomes more consequential when the machine must react quickly to a moving object, a collision risk or an unexpected change in its environment.

Cloud operation can offer access to more computing power and make model updates easier to distribute. However, it also ties physical behavior to network performance. Industrial customers may ultimately want systems that combine cloud-based reasoning with local controls, allowing routine movements and emergency responses to continue even when connectivity is limited. The available sources do not establish whether Gemini Robotics 2.5 uses that kind of hybrid arrangement.

DeepMind also introduced ASIMOV-Jailbreak, a benchmark intended to test whether embodied models reject unsafe instructions. The ASIMOV Agentic Safety Evaluation dataset describes the benchmark as a physical-safety and multi-turn adversarial-jailbreak suite. Its datasets and scripts are designed to evaluate unsafe instructions in robotics scenarios, including situations in which a user attempts to manipulate a model across multiple exchanges.

DeepMind reported a 99.5% jailbreak-resistance result for Gemini Robotics 2.5. That number is useful because ordinary task success does not measure whether a robot knows when not to act. A machine that follows instructions accurately but accepts dangerous commands could be more hazardous than one that performs fewer tasks.

Yet safety benchmarks have the same limitation as dexterity benchmarks: they provide evidence, not a complete deployment guarantee. Refusing unsafe instructions in a test suite does not prove that a robot will handle every ambiguous command, unusual object or rapidly changing factory condition correctly. The benchmark is an important attempt to formalize that problem, but its practical value will depend on how closely its scenarios match real operations.

The supplied reception material contains only Whitwam’s critical assessment. No opposing commentary was provided from Mercedes-Benz, Apptronik or another independent evaluator. That means the strongest available evidence remains Google’s own performance claims, alongside the reported concern that 96% is not yet close enough to industrial dependability.

The broader implication is that robotics competition is shifting. The challenge is no longer only to make machines move, grasp and generalize across tasks. It is to make them reliable enough that human supervisors are not simply correcting a new class of errors. Gemini Robotics 2.5 shows Google narrowing that gap, but the last 4% may determine whether these systems become flexible factory tools or remain impressive demonstrations with expensive human oversight.

#Gemini Robotics 2.5#Google DeepMind#Apptronik#Apollo 2#Mercedes-Benz#Ryan Whitwam#ASIMOV-Jailbreak

Alex Carter is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.

The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.

How a story is made

  • Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
  • Writing. The writing model drafts it against what was found, at Alex Carter's usual length and in Alex Carter's usual register.
  • The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
  • Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
  • Fact check. A last pass hunts for claims the article makes and its sources do not.
  • A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.

If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.

This article was generated using AI and published automatically without human pre-publication review.

How this article was made

The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.