What happened
Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, describing it as its most capable vision-language-action model to date. The model is designed to expand the breadth of skills robots can learn and the speed at which they acquire them. The announcement says the team evaluated it across more than 200 distinct tasks, including serving food with tongs, tying a bow, and pipetting liquid. These tasks combine tool use, manipulation of deformable materials, and handling delicate objects.
Against previous baselines, Google reports task success rates 30% to 100% higher, depending on task type, and completion times 20% to 30% faster. The model is intended to work with different robot types. Examples named include Apptronik’s Apollo 2 humanoid, with a five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-finger parallel grippers. Google frames the work as a step toward robots that can operate in unstructured environments rather than only in tightly controlled demonstrations.
For a workplace test, Google partnered with Apptronik and Mercedes-Benz on active automotive production lines in Kecskemet, Hungary, and Sindelfingen, Germany. Apollo 2 performed automotive kitting: taking metal brackets from bulk supply bins and placing them into sequencing containers in the order required for vehicles on the line. Google says the system used Gemini Robotics ER to track the workflow and Gemini Robotics 2.5 to control the robot’s hands and arms. The company reports a 96% success rate in this proof-of-concept and says the models improved in a few weeks of post-training, with little drop-off in the Mercedes-Benz production setting.
The announcement also describes safety work. Inference-time reasoning guidance steers the robot around physical hazards. Google reports that this improved safe task completion by 33% around obstacles and 43% near people in its evaluations. It also announced ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts disguising unsafe requests across dozens of scenes. Google says Gemini 4 Argon achieved 99.5% jailbreak resistance on it, compared with 82% for the previous Gemini Robotics ER 2 baseline. Gemini Robotics 2.5 is rolling out to robotics partners.
What the archive adds
New Runtime’s earlier coverage described Gemini Robotics ER 2 entering public preview on Google Cloud on October 27, 2026 (earlier coverage). This is a new event in the same robotics family: the announcement focuses on the action model and reports industrial task results, rather than the earlier preview of an embodied reasoning model.
Trend signal
This supports the pattern of agent economics moving to completed work: the headline evidence is a success rate on physical tasks, not a token price or model-call count. The factory pilot makes task completion the relevant unit, although a single proof-of-concept does not establish production economics or repeatability across sites.
Why now
Google says Gemini 4 Argon accelerated robotics research by generating simulated practice runs from real camera scenes and helping researchers analyze failures. The company frames that capability as a development accelerator for Gemini Robotics 2.5: simulated practice, real demonstrations, and failure analysis are presented as a loop for adapting robot behaviour to new scenes and objects.
Connections to other events
The release follows the October 27 public preview of Gemini Robotics ER 2 on Google Cloud (earlier coverage). The two releases describe complementary roles: ER 2 reasons about physical environments and coordinates tasks, while Gemini Robotics 2.5 is the action model controlling hands and arms in the Mercedes-Benz kitting pilot.
Why it matters
For teams considering robots in structured industrial workflows, the announcement offers a more operationally relevant signal than a dexterity demonstration alone: kitting on an active production line, with reported task success and a defined coordination between a reasoning model and an action model. If the result generalizes, robots may take on repetitive preparation work while technicians concentrate on more complex assembly tasks.
The evidence is not yet enough to plan a deployment. A 96% task success rate still implies failures that need detection, recovery, and human ownership, and the post-training result comes from one described pilot. The safety improvements are also evaluation results, not a guarantee that a robot is safe in every factory or around every worker. Builders should treat the announcement as a signal to request task-level reliability, recovery, and safety data before comparing it with their own process requirements.
The practical architectural point is the division of responsibilities. Google describes Gemini Robotics ER as maintaining the workflow while the 2.5 action model controls manipulation. A team evaluating physical automation should similarly distinguish process coordination, sensing, action execution, exception handling, and the authority to resume work after a failure. A model improvement does not remove the need for those operational boundaries.
Limits and uncertainty
The results are reported by Google DeepMind and are not independently verified. The 96% result comes from a proof-of-concept on specified Mercedes-Benz lines, so it should not be read as a general rate for factories, shifts, parts, or robot fleets. Google reports relative evaluation gains and safety-task improvements, but those figures do not by themselves measure operating cost, throughput, uptime, intervention burden, or the consequences of a failed action. Partner rollout establishes an access path, not broad availability or a deployment commitment.
Skill to build now
Turn one candidate physical workflow into an evaluation contract this week. Define the objects, starting states, ordered outputs, acceptable error types, recovery procedure, human escalation point, and safety stop condition. Measure successful completion separately from whether a human had to intervene. That makes a vendor-reported success rate comparable with the work your team actually needs a robot to perform, and exposes whether the bottleneck is reasoning, grasping, sensing, or process recovery.
Reader takeaway
Gemini Robotics 2.5 is a meaningful step from lab dexterity tests toward factory tasks, with reported gains in completion and safety. But one partner proof-of-concept is not deployment evidence at scale: teams should ask for repeated, task-specific reliability, recovery, and intervention measurements before treating the reported success rate as an operational forecast.