Google DeepMind's Gemini Robotics 2.5 Hits 96% in Mercedes Factory Pilot

Google DeepMind has released Gemini Robotics 2.5 , a vision-language-action model for controlling robots from images and natural-language instructions. DeepMind reports higher task success, faster execution, and a 96% success rate in an automotive kitting pilot with Apptronik and Mercedes-Benz. Access is limited to existing robotics partners, with no public API or downloadable model announced.

A vision-language-action model, commonly shortened to VLA, maps camera frames, robot state, and an instruction to motor commands. This approach reduces dependence on separately engineered perception, planning, and control systems, although deployments still require hardware integration, safety controls, and task-specific evaluation.

A 200-task manipulation test

DeepMind evaluated the model on 200 tasks spanning tabletop pick-and-place, multi-step contact work, precision insertion, tool use, and deformable objects. Examples include serving food with tongs, tying a bow, folding clothes, and pipetting liquid.

Across the published benchmarks, DeepMind reports success-rate gains of 30% to 100% over earlier baselines and completion-time reductions of 20% to 30%. Success rates depend on trial counts, reset policies, scene variation, and human intervention, so the task-level figures provide more context than the aggregate ranges.

Tasks involving fabric, ribbons, food, and hand tools remain difficult because object shape changes during manipulation and small errors compound over several actions. The reported gains indicate better handling of those conditions, while the 40% results for packing and folding also show substantial room for improvement.

Argon turns scenes into training data

DeepMind says it used Gemini 4 Argon to generate simulated practice runs from real camera scenes. Argon constructs a digital representation of the scene, varies objects and conditions, and produces synthetic trajectories that are mixed with demonstrations collected through human teleoperation.

Argon also analyzes failed rollouts by reviewing video, proposing possible causes, and testing changes in simulation or on a physical robot. This process can shorten the cycle between identifying a failure mode and collecting targeted training data.

The release does not quantify the ratio of synthetic to real demonstrations or the amount of platform-specific fine-tuning. Those details determine the cost of reproducing the approach and the extent to which simulation replaces physical data collection.

Two models run the factory task

Mercedes-Benz and Apptronik tested the system at automotive facilities in Kecskemét, Hungary, and Sindelfingen, Germany. The robots performed kitting, which requires selecting parts from bulk bins and arranging them in trays according to the build sequence for each vehicle.

The pilot divides control between two models. Gemini Robotics ER, an embodied-reasoning model, tracks the workflow and selects actions over longer time horizons. Gemini Robotics 2.5 controls the arms and hands as the robot reaches into bins, grasps metal brackets, and places them into sequencing containers.

DeepMind reports a 96% success rate for the pilot. Comparing that figure with other deployments requires the number of attempts, run duration, human-intervention rate, failure criteria, and reset policy, which the release does not specify.

DeepMind also describes the action model as hardware-agnostic. The same checkpoint reportedly operates Apptronik’s Apollo 2 , equipped with a five-finger SharpaWave hand with 22 degrees of freedom, and the Franka Duo, which uses two-finger parallel grippers. Each platform still requires calibration, action-space mapping, and safety integration; the published material does not quantify that engineering work.

Safety spans motion and prompts

DeepMind evaluates two classes of risk: physical collisions during movement and adversarial instructions delivered through the language interface. The model uses inference-time reasoning guidance to revise motions around detected hazards, with reported safe-completion gains of 33% near obstacles and 43% near people.

For language-level attacks, DeepMind released ASIMOV-Jailbreak , a public dataset of prompts that disguise unsafe requests across multiple scenes. Argon achieved a reported jailbreak-resistance rate of 99.5%, compared with 82% for the Gemini Robotics ER 2 baseline, a difference of 17.5 percentage points.

Physical and language evaluations cover different failure paths and should remain separate in test suites. A model can reject an unsafe instruction while still producing a hazardous trajectory, or accept a benign instruction and collide because its perception or control fails.

The engineering takeaways

The results show broader manipulation capability under DeepMind’s test conditions and a path from laboratory evaluation to a production-site pilot. They do not establish unattended operation across arbitrary factories, hardware, or tasks, and the missing deployment details limit cost and reliability comparisons.

• Data: Generated simulation can supplement teleoperation demonstrations, especially for rare failures and scene variation.
• Architecture: Separating long-horizon reasoning from high-frequency motor control gives each model a narrower responsibility.
• Portability: A shared checkpoint may reduce retraining across robot platforms, although adapters and calibration remain necessary.
• Evaluation: Test suites need physical-hazard scenarios, task reliability metrics, and language-level jailbreak attempts.
• Deployment: Trial counts, intervention rates, reset policies, latency, and hardware requirements are necessary for comparing systems.

DeepMind is distributing Gemini Robotics 2.5 through its existing partner program and has disclosed neither public API access nor pricing. Developers without partner access can use the ASIMOV-Jailbreak dataset, while model weights, inference endpoints, and training code remain unavailable.