JOURNAL / 2026.11.12
Gemini Robotics 2.5 improves manipulation and brings Apollo 2 to an industrial trial
DeepMind introduces a new robot controller with gains in dexterity and a Mercedes-Benz trial. The result opens possibilities for flexible automation, but assessing sustained work and error recovery still requires more evidence.
Preparing the parts a vehicle needs seems like a modest task beside the promises of a general-purpose robot. That is precisely what makes it an interesting test: there is a recipient, an order to follow, and a consequence when the delivery goes wrong.
DeepMind introduced Gemini Robotics 2.5 on November 10. It reports manipulation improvements across more than 200 tasks and 96% success preparing components with Apptronik's Apollo 2 on Mercedes-Benz production lines. This is a proof of concept reported by the provider. The announcement does not specify the number of attempts, duration, or exact unit behind that percentage.
The change deserves attention without crediting it with everything that already existed. In July, Gemini Robotics 2 already offered whole-body control, multifinger hands, and collaboration between robots. That publication acknowledged that multifinger manipulation remained difficult. The current advance develops a capability that can determine whether an agent's plan gets carried out: grasping, placing, and handling objects with sufficient dexterity.
The family documentation helps locate the advance. The embodied reasoning model, ER, organizes steps; the vision-language-action model, VLA, turns observations and instructions into motor control. Improving that second component can expand the tasks a system handles where simply changing the wording of its instructions would not suffice. For someone designing a physical application, the question is which movements and recoveries they can entrust to the controller, alongside which goals they can describe to it.
Work continues after an unsuccessful attempt
A success rate leaves an engineering decision unresolved: what happens when an operation goes wrong? Imagine a part left tilted in its slot. The robot could correct it, ask for help, or mark it as placed. Those outcomes have different consequences for the next station. This is our own example; it does not describe an observed incident in the pilot.
I would evaluate an installation by recording accepted complete deliveries, cycle times, and human interventions, with an explicit definition of each failure. I would also separate retry time from the time needed to recover a situation requiring a stop. A system that reliably detects its limits could be more useful than one with better initial success and silent errors. Without those data, I would not estimate how many working hours it saves or which investment pays off.
Safety adds another layer. The official ASIMOV repository incorporates ASIMOV-Jailbreak to evaluate adversarial requests and decisions about physical actions. It supports studying whether a system accepts a dangerous instruction; that does not certify that an allowed trajectory avoids a collision. Both questions belong in an evaluation, alongside physical checks outside the agent's reasoning. Publishing evaluation material enables scrutiny, although accessing its files requires accepting access conditions.
Demonstration and availability also need to be distinguished. The dedicated Robotics 2.5 page points to an early-access waitlist. A team should not plan an integration as though the controller were already available for any hardware. It would first need to confirm access, compatibility, adaptation, and operating conditions.
My reading is that the advance warrants bounded industrial trials with acceptance criteria agreed with the people receiving the work. Flexibility will be valuable if it reduces the effort of changing tasks without shifting more supervision onto staff. I would measure that tradeoff directly: what teaching a variant costs, how many deliveries remain correct, and who resolves exceptions. That evidence would help determine where a more capable controller becomes an everyday tool.
Primary sources consulted on November 12, 2026. We have not tested the model or audited the pilot. Examples and evaluation proposals are our own analysis.