Google DeepMind has tested its Gemini Robotics 2.5 artificial intelligence model on active Mercedes-Benz production lines, enabling Apptronik's Apollo 2 humanoid robots to sort automotive components with a reported task success rate of 96%.
The proof of concept took place at Mercedes-Benz manufacturing facilities in Sindelfingen, Germany, and Kecskemét, Hungary. The robots performed parts kitting, retrieving metal brackets from bulk containers and arranging them in the sequence required for individual vehicles.
Announced on 10 November, the trial forms part of Google DeepMind's introduction of Gemini Robotics 2.5, a vision-language-action model developed to improve robotic manipulation, adaptability and performance across different physical tasks.
The system combines visual interpretation with action generation, allowing robotic equipment to respond to the changing position of components rather than depending entirely on a predetermined movement sequence.
In the Mercedes-Benz application, Apollo 2 retrieves individual brackets from supply bins and transfers them into sequencing containers. The operation requires the robot to recognise the required component, identify a suitable grasp and position the part according to the manufacturing sequence.
Parts kitting is used in automotive manufacturing to prepare sets of components before they reach an assembly operation. Rather than collecting individual parts from multiple storage locations during assembly, a production worker or automated system can receive a prepared selection matched to the vehicle being manufactured.
The method depends on the correct components being selected and presented in the required order. A picking error can introduce a missing or incorrect item into the sequence, potentially requiring intervention before subsequent production operations proceed.
Automating the activity becomes more demanding when components are stored loosely in bulk containers. Parts can overlap, obscure one another or change orientation as the bin is emptied, requiring the handling system to respond to conditions that differ between successive movements.
Conventional robotic picking can address these applications using appropriate machine vision, gripping equipment and programmed control. The engineering requirements increase when component positions vary substantially or when several different parts must be recognised and handled within the same operation.
Gemini Robotics 2.5 is designed to support more adaptable manipulation by interpreting sensor information and generating movements suitable for the current task. It operates alongside an embodied reasoning model that maintains information about the wider sequence of actions.
In the Mercedes-Benz trial, Gemini Robotics ER was used to track the required workflow, while Gemini Robotics 2.5 controlled the movements of the robot's arms and hands.
The reasoning component helps establish which operation should be performed next. The action model then generates movements to retrieve and position the component, using updated information as the arrangement of parts changes.
This division of functions supports tasks requiring several connected decisions. A robot may need to establish which part is required, identify where it can be grasped, move it clear of adjacent components and confirm its position within the completed kit.
Apptronik's Apollo 2 platform incorporates articulated arms and five-fingered hands. The demonstrated configuration uses the SharpaWave hand, which provides 22 degrees of freedom for manipulating objects.
These degrees of freedom describe the independent movements available within the mechanical hand. They do not establish the weight of an object it can handle, the force available at each finger or the speed at which a completed manipulation can be performed.
Grasping performance also depends on the object's geometry, surface condition and position. A component with a suitable accessible edge may be easier to retrieve than one lying beneath other parts or presenting limited contact surfaces.
Google DeepMind reported 96% success during the automotive kitting proof of concept. The company has not disclosed the number of attempts represented by that percentage, the precise success criteria or the frequency of human intervention.
The result consequently describes reported task performance under the trial conditions rather than the availability or throughput of a fully deployed production system.
Assessing sustained manufacturing performance would require additional information, including the number of kits completed, processing time, incorrect selections, recovery procedures and the proportion of operating time during which the system remained available.
A successful individual pick does not necessarily mean that a complete kit has been prepared within the time allowed by an assembly process. The required production rate depends on the number of components in each kit, their presentation and the time needed to complete successive movements.
Robot manipulation is also constrained by the physical equipment. The software must generate actions that can be executed within the platform's mechanical reach, available joint movement, gripping capability and stability limits.
Environmental conditions can affect perception. Changes in lighting, reflections from metal components or obstruction of the camera's view may alter the information available to the model, requiring appropriate handling of uncertainty during operation.
Google DeepMind evaluated Gemini Robotics 2.5 across more than 200 tasks involving different forms of manipulation. The company reported improvements of between 30% and 100% in task success rates against selected earlier model baselines, with completion times reduced by approximately 20% to 30%.
Those comparisons relate to Google's broader evaluations, which included activities beyond industrial manufacturing. They should not be interpreted as equivalent improvements in productivity or cycle time at the Mercedes-Benz facilities.
The model has also been demonstrated on different robotic platforms, including the Franka Duo system with parallel grippers. Supporting different mechanical arrangements requires the software to account for the movements and interfaces available on each robot.
Google is using Gemini 4 Argon to support the development of these capabilities through simulation and analysis of task failures. Simulated environments allow researchers to introduce variations in object position and operating conditions before conducting further physical tests.
Training against simulated scenarios can improve the range of situations considered during development, although differences between a simulation and the physical environment must still be addressed through testing on actual equipment.
Safety is being developed alongside manipulation performance. Google reported improvements in safe task completion during evaluations involving obstacles and people, using additional reasoning guidance intended to influence robot behaviour.
These software evaluations are distinct from the safety assessment of a complete industrial installation. Robotic equipment operating near people must be engineered and assessed according to its physical hazards, operating conditions, protective measures and applicable requirements.
Google DeepMind is making Gemini Robotics 2.5 available to selected robotics partners. The Mercedes-Benz application remains a proof of concept, with no commercial deployment quantity or timetable for a factory-wide rollout confirmed.
Further industrial testing will establish whether the reported kitting performance can be maintained across longer operating periods, different component presentations and the production schedules used at Mercedes-Benz's manufacturing facilities.
