On 10 November 2026, Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model that converts what a robot sees and hears into physical movement. The announcement matters not because it is the first robotics model, but because it is the first to show production-grade reliability on an active automotive factory floor. During a proof-of-concept at Mercedes-Benz, humanoid robots powered by Gemini Robotics 2.5 achieved a 96% success rate sorting automotive parts into sequencing containers on live production lines in Hungary and Germany.
For enterprise buyers, the implication is direct. Until now, agentic AI has been a software phenomenon: chatbots, coding assistants, data analysts and workflow automators operating entirely inside screens and servers. Gemini Robotics 2.5 is a signal that the same agentic paradigm is crossing into the physical world, with safety benchmarks, factory-floor metrics and a partner programme already underway.
What Gemini Robotics 2.5 does
The model is a vision-language-action architecture, which means it processes visual and linguistic input and outputs motor control commands directly. This is different from the traditional pipeline in which a language model reasons, a vision model interprets, and a separate controller moves the hardware. By unifying perception, reasoning and action in a single model, Gemini Robotics 2.5 reduces the latency and failure points that have historically kept robots in carefully scripted environments.
DeepMind evaluated the model across more than 200 distinct tasks, from serving food with tongs to tying a bow on a gift box to pipetting liquid in a lab setting. Compared with previous baselines, success rates improved by between 30% and 100% depending on task type, while completion times dropped by 20% to 30%. The model works across robot types, including Apptronik's Apollo 2 humanoid with a five-finger, 22-degree-of-freedom hand, and the Franka Duo industrial platform with parallel grippers.
The breadth matters as much as the accuracy. Industrial robotics has historically been narrow: one robot, one task, one environment. Gemini Robotics 2.5 is designed for generalisation, the ability to handle tasks it was not explicitly trained on by reasoning about the physical situation in real time. That shift from narrow automation to general physical intelligence is the same shift that happened in software over the past three years, and it is now arriving in hardware.
The Mercedes-Benz proof-of-concept
The most important number in the announcement is 96%. DeepMind partnered with Apptronik and Mercedes-Benz to deploy Apollo 2 humanoid robots on active production lines in Kecskemet, Hungary, and Sindelfingen, Germany. The task was automotive kitting: sorting individual parts from bulk bins into organised trays in the exact sequence required for each vehicle on the assembly line.
This is not a laboratory demo. Kitting is a repetitive, physically demanding task that keeps assembly lines moving. Getting it wrong stops production. The 96% success rate was achieved not in simulation but on live factory floors, using the Gemini Robotics ER embodied reasoning model to track workflow state and Gemini Robotics 2.5 to control the robot's arms and hands.
For Mercedes-Benz, the business case is labour reallocation rather than labour replacement. By offloading monotonous kitting work to robots, technicians can focus on complex build tasks that require human judgement. The model is the same one that software agent vendors have been selling for two years: the agent handles the routine, the human handles the exception.
Safety by design, not by bolt-on
Physical agents carry risks that software agents do not. A coding agent that hallucinates produces bad code. A robot that hallucinates can collide with a person. DeepMind has addressed this with three layers of safety architecture.
Inference-time reasoning guidance. The model steers around physical hazards in real time, much like a safety classifier filters harmful outputs in a language model. In evaluations, this guidance increased safe task completion around obstacles by 33% and around nearby people by 43%.
Adversarial jailbreak testing. DeepMind released ASIMOV-Jailbreak, an open-source benchmark containing hundreds of prompts designed to trick robots into unsafe actions across dozens of physical scenes. Gemini 4 Argon, which powers the reasoning layer, achieved a 99.5% jailbreak resistance rate. For context, the previous baseline scored 82%.
Multi-robot coordination. The embodied reasoning model supports multi-robot collaboration, allowing different robots to divide labour and communicate state. In a factory setting, this prevents collisions between autonomous units and enables coordinated workflows that no single robot could complete alone.
Accelerated by Gemini 4 Argon
Behind the robotics model is a research acceleration technique that may be just as significant for enterprise buyers. DeepMind used Gemini 4 Argon as a simulation engine to generate thousands of varied training runs from a single real-world camera scene, building realistic digital twins that let the robot generalise faster than training on physical demonstrations alone.
The technique solves one of the hardest problems in robotics deployment: data scarcity. Industrial robots traditionally require thousands of physical trials before they can handle a new part or environment. By using a frontier intelligence model to simulate those trials, DeepMind cut the time from camera capture to factory deployment dramatically. For enterprises, this means the cost and timeline of deploying physical agents is starting to look more like software agents.
What this means for enterprise buyers
For procurement and operations teams evaluating agentic AI, the Gemini Robotics 2.5 announcement carries four practical signals.
The agentic boundary is expanding. Enterprises that have been planning agentic AI for software workflows should now include physical workflows in their evaluation. Warehousing, manufacturing, logistics and facilities management are all entering the agentic domain. The same governance, observability and safety questions that apply to software agents now apply to robots.
Generalisation is the new metric. Traditional robotics procurement focuses on speed, payload and repeatability for a single, known task. Agentic robotics procurement should focus on adaptability: can the robot handle a part it has not seen before, respond to a change in the environment, and escalate to a human when uncertain? The 200-task evaluation set is a benchmark buyers should ask vendors to replicate.
Safety must be verifiable, not assumed. DeepMind's release of ASIMOV-Jailbreak and its public adversarial testing results set a standard that enterprise buyers should expect from any physical agent vendor. Ask for jailbreak resistance rates, collision avoidance test data, and evidence of human-aware navigation. If a vendor cannot provide them, they are not ready for shared workspaces.
The data platform matters for physical agents too. Just as software agents need access to enterprise data to be useful, physical agents need access to factory data, inventory systems and production schedules. The integration between the robot's reasoning layer and the enterprise's operational systems is where many deployments will succeed or fail. Buyers should evaluate the API and data connectivity layers before evaluating the hardware.
What this means for suppliers
For robotics and AI vendors, the announcement resets competitive expectations in three ways.
Software and hardware are converging. Apptronik is a robotics hardware company. DeepMind is an AI research lab. Their partnership on a production factory floor shows that the winning products will be integrated systems, not standalone robots or standalone models. Vendors selling hardware without intelligence, or intelligence without a hardware path, will face pressure to partner or acquire.
Simulation is becoming a core competency. The ability to train robots in simulated environments before deploying them physically is no longer a research curiosity. It is a cost and speed advantage. Vendors that can demonstrate rapid sim-to-real transfer will outcompete those still relying purely on physical training.
Enterprise buyers will ask for agentic governance on the factory floor. The same CISOs who are demanding inventory, observability and audit trails for software agents will demand them for physical agents. Vendors should anticipate questions about who authorised a robot's actions, what it did, and whether it stayed within its permitted parameters.
The Agentic Expo angle
Agentic Expo, at Olympia London on 23 to 24 March 2027, brings together the buyers and suppliers navigating this exact transition. The attendees deciding which agent platforms to trust in 2027 will not be choosing between software agents and physical agents. They will be choosing platforms that can govern both.
The Mercedes-Benz proof-of-concept is a preview of what enterprise agentic AI looks like when it stops being a pilot and starts being an operating model. The robot sorts parts. The technician builds vehicles. The governance layer tracks what happened. That is the architecture that will define the next phase of enterprise automation, and it is the architecture that will be on display at Olympia in March.
Sources: Google DeepMind, "Gemini Robotics 2.5 Accelerates Real-World General Dexterity", 10 November 2026; Google DeepMind, "Gemini Robotics", November 2026.