Table of Contents
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model that expands the range of tasks robots can perform and improves their speed and dexterity. In a proof of concept on active Mercedes-Benz production lines, Apptronik’s Apollo 2 humanoid used DeepMind’s models to pick and organize automotive parts with a reported 96% success rate.
The release brings together improvements in physical manipulation, safety guidance, and training methods. DeepMind evaluated the model across more than 200 distinct tasks and is rolling it out to its robotics partners.
In its launch demonstration, DeepMind showed the robot folding laundry, serving pastries, and sorting car parts. The factory pilot applies that broader skill set to a specific industrial problem, preparing the right components in the right order for vehicle assembly.
How Apollo 2 Handles Automotive Kitting
The Mercedes-Benz proof of concept took place in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 autonomously performed automotive kitting, taking metal brackets from bulk supply bins and placing them into sequencing containers in the order needed for each vehicle.
That work involves more than moving an object between two fixed locations. The robot has to pick parts from a bulk bin, track which positions need filling, and correct gaps as it prepares the sequence for assembly.
DeepMind divides those responsibilities between two models. Gemini Robotics ER, its embodied reasoning model, keeps track of the workflow. Gemini Robotics 2.5 controls Apollo’s hands and arms to carry out the physical actions. The reasoning layer manages what needs to happen; the action model handles the movements needed to execute it.
DeepMind robotics head Carolina Parada reported the 96% result as accuracy in picking and organizing parts for assembly. The figure applies to this automotive-kitting proof of concept, rather than the model’s performance across its full range of tasks.
Apptronik CEO Jeff Cardenas said the models improved substantially within a few weeks of post-training and transferred to Mercedes-Benz production with very little performance drop-off. Post-training adapts an already trained model to particular tasks or operating requirements, giving the partners a way to refine general manipulation capabilities for factory work.
Broader Dexterity and Faster Training
Beyond the factory pilot, DeepMind tested skills such as serving food with tongs, tying a bow on a gift box, and pipetting liquid. These tasks combine tool use, delicate handling, and manipulation of flexible objects whose shapes change as the robot works.
In DeepMind’s laboratory evaluations, Gemini Robotics 2.5 achieved success rates 30% to 100% higher than previous baselines, depending on the task type, while completing tasks 20% to 30% faster. Those are relative improvements over the baselines, separate from the 96% factory-pilot result.
The model supports different robot configurations. Apollo 2 uses five-finger SharpaWave hands with 22 degrees of freedom, while the Franka Duo platform uses two-finger parallel grippers. Degrees of freedom describe the independent movements available to a mechanism; the two platforms give the model substantially different ways to grasp and manipulate objects. Apptronik’s demonstration post pointed to both bow-tying and car-part sorting as examples of Apollo’s range.
Gemini 4 Argon also contributed to developing the new action model. It generates varied simulated practice runs from real-world camera scenes, builds digital twins, and helps researchers analyze failures from video. DeepMind combines that synthetic experience with real training demonstrations so the robot can learn to handle different objects and environments.
Google’s Argon robotics research examples show how that process works. A simulation can be calibrated against recorded robot footage, including the timing and magnitude of fingertip contact forces, rather than simply resembling the scene visually. Researchers can then test changes in simulation before trying them on hardware.
Argon’s role in that research process is distinct from the model controlling the robot’s immediate movements. Ars Technica reports that Gemini Robotics 2.5 currently runs on cloud servers, making network and processing latency part of the operating setup.
Safety Guidance and Partner Access
Gemini Robotics 2.5 adds inference-time reasoning guidance to steer robot actions around physical hazards. Inference is the stage when a trained model processes new information and decides how to respond, allowing guidance to influence behavior while a task is underway.
DeepMind reports that this capability improved safe task completion by 33% around obstacles and by 43% around nearby people in its evaluations. The guidance can help a robot pause or wait when its working area is obstructed, rather than continuing a movement through a hazard.
The company also released ASIMOV-Jailbreak, a benchmark for testing whether embodied agents can resist requests that disguise unsafe actions as ordinary tasks. Its evaluation suite includes paired harmful and benign tabletop instructions, alongside 845 multi-turn adversarial episodes that can run for up to 10 steps.
Testing benign requests alongside harmful ones lets researchers examine both unsafe compliance and unnecessary refusal of legitimate work. The repository provides evaluation scripts and scoring tools, although access to its files requires accepting the Hugging Face access conditions.
DeepMind is distributing Gemini Robotics 2.5 through its robotics partnerships rather than a general public release. The rollout gives partners access to the new manipulation model for testing and integration on their own hardware and tasks.
