One-sentence summary
On November 10, 2026, Google DeepMind released Gemini Robotics 2.5, its most capable vision-language-action model to date. On active Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 used the model to sort automotive parts with a 96% success rate; the company also reported a 30% to 100% higher success rate than prior baselines and 20% to 30% faster task completion.
Event overview
On November 10, 2026, Google DeepMind published Gemini Robotics 2.5 on its official blog, describing it as the company's most capable vision-language-action (VLA) model to date. The post was written by Carolina Parada, who leads the Gemini Robotics team, and appeared in the Models section of the DeepMind blog.
The test setting was not a lab demo. DeepMind says it worked with Apptronik and Mercedes-Benz to run Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task was automotive kitting: pulling metal brackets from bulk bins and placing them into sequencing trays in the exact order required for each vehicle on the line.
DeepMind reports a 96% success rate for Apollo 2 performing this kitting task autonomously in the proof-of-concept. The robot used Gemini Robotics ER to track the workflow and the Gemini Robotics 2.5 action model to control its hands and arms. When it encountered gaps in a supply bin, the robot corrected its trajectory rather than repeating a fixed motion.
This is not the first Gemini Robotics release. Earlier in 2026, DeepMind introduced Gemini Robotics 2 with a focus on whole-body intelligence across many robot types. Gemini Robotics 2.5 shifts the emphasis from whole-body control to learning a wider range of tasks faster, and puts real production-line reliability at the center of the launch.
Key facts
| Item | What Google DeepMind disclosed |
|---|---|
| Release date | November 10, 2026 (official DeepMind blog) |
| Model | Gemini Robotics 2.5, a vision-language-action model |
| Evaluation scale | More than 200 distinct tasks |
| Task examples | Using tools, handling deformable objects and delicate items; serving food with tongs, tying a bow, pipetting |
| Production sites | Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany |
| Robot | Apptronik Apollo 2, with a five-finger, 22-DOF SharpaWave hand |
| Line task | Automotive kitting: metal brackets from bulk bins into sequencing trays |
| Measured success | 96% in the proof-of-concept |
| Task improvement | 30% to 100% higher success rate and 20% to 30% faster completion versus prior baselines, by task type |
| Safety improvement | 33% higher safe task completion around obstacles; 43% higher around nearby people |
| Jailbreak resistance | Gemini 4 Argon reached 99.5% on the ASIMOV-Jailbreak benchmark, versus 82% for Gemini Robotics ER 2 |
| Availability | Rolling out to robotics partners; no mass-production deployment date announced |
| Training accelerator | Gemini 4 Argon acts as a simulation engine, generating thousands of simulated practice runs from real camera scenes |
What the 96% actually measured
The 96% figure applies to automotive kitting, not to robot manipulation in general. The task chain is specific: Apollo 2 takes metal automotive brackets from bulk bins and places them into containers arranged in the order each vehicle needs. Parts are not fixed in place inside the bin, so the robot has to judge orientation before grasping and correct position errors during placement.
Two models split the work. Gemini Robotics ER understands the workflow, tracks steps and decides what comes next; the Gemini Robotics 2.5 action model turns those decisions into hand and arm movements for grasping and placing. DeepMind stresses that the robot corrected for gaps in the supply bin instead of replaying a fixed trajectory.
The 96% therefore measures the success rate of autonomous kitting inside a proof-of-concept. It is not a factory takt-time achievement rate, overall equipment effectiveness (OEE) or continuous uptime. Google has not disclosed how many shifts the test covered, how many parts were handled, how often humans intervened, or the cycle time per part. Those metrics determine whether the result can move into regular production.
Why kitting is hard for a VLA model
Automotive kitting is a low-value, highly repetitive, sequence-critical station. Parts move from bulk bins to sequencing trays, and three sources of difficulty show up: random part orientation, gaps inside the bin, and a different build sequence for each vehicle. A traditional fixed-trajectory robot needs fixtures and paths for every part and every bin; a VLA model attempts to generalize to new objects from visual and language input.
DeepMind's evidence for generalization is the set of more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. The company says Gemini Robotics 2.5 achieved a 30% to 100% higher success rate than prior baselines and completed tasks 20% to 30% faster. Those figures are ranges across task types, not an average for a single task.
On the hardware side, Apollo 2 uses a five-finger, 22-degree-of-freedom SharpaWave hand. The model also supports the Franka Duo dual-arm platform with two-fingered parallel grippers. DeepMind says the same model natively works with different robot embodiments without training a separate action policy for each robot.
Jeff Cardenas, co-founder and CEO of Apptronik, said in the official post that Apollo 2 improved dramatically in just a few weeks of post-training with Gemini Robotics, and carried over to Mercedes-Benz production with very little drop-off. That statement describes post-training efficiency and transfer performance, not long-term line stability.
Safety: physical steering and jailbreak resistance
Gemini Robotics 2.5 reports two sets of safety numbers. The first covers inference-time reasoning guidance: the robot steers around physical hazards in real time while executing a task. DeepMind says this capability raised safe task completion by 33% around obstacles and by 43% around nearby people.
The second covers agentic jailbreak resistance. DeepMind released ASIMOV-Jailbreak, an open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes. Gemini 4 Argon reached a 99.5% jailbreak resistance rate on that benchmark, compared with 82% for the previous Gemini Robotics ER 2 baseline.
The two numbers address different risks. Physical safety is about not hitting people or objects in a messy environment; instruction safety is about refusing a malicious prompt that tries to make the robot perform a dangerous action. DeepMind released both in the same model generation, extending safety evaluation from single actions to long-horizon agentic tasks.
The real limits of humanoids on the line
First, 96% is proof-of-concept data. The Mercedes-Benz plants are active lines, but the test remains a pilot: Google does not say Apollo 2 has been integrated into regular takt time, nor does it give the number of deployed robots, operating hours or parts handled per shift. A 96% success rate demonstrates task feasibility, not volume reliability.
Second, cycle time and cost are undisclosed. Automotive lines organize work around a time window per vehicle, and the robot has to finish kitting without slowing the main line. Google has not published per-part cycle time, changeover time, maintenance intervals or unit cost, so payback cannot be calculated. This site does not estimate it.
Third, generalization has boundaries. The 200-plus tasks cover tool use, deformable objects and delicate handling, but Google does not give success rates per task category or a failure-mode distribution. The 96% line result applies to one object family, automotive brackets; other parts, bins or plants require new post-training and validation.
Fourth, the safety benchmark is an open evaluation, not a site certification. ASIMOV-Jailbreak tests adversarial prompts, and the 33% and 43% physical-safety gains come from Google's evaluation environment. Converting those results into factory safety certification depends on regional machine-safety standards and on-site risk assessment.
Fifth, the deployment timeline is open. DeepMind says Gemini Robotics 2.5 is rolling out to robotics partners but has not announced a mass-production deployment date. For Mercedes-Benz and other manufacturers, the confirmed fact is a completed proof-of-concept on active lines, not a fleet already on the job.
Impact and limits
The immediate impact is a public, quantified data point for VLA models in automotive production. Until now, most humanoid demonstrations were lab videos; DeepMind has put a success rate, two active plants and a named task on the record. That raises the bar for competitors and gives manufacturers a concrete reference for pilot design.
The second impact is the shift from "can the robot do it" to "can the line absorb it". Kitting is only one station. A production deployment also needs safety certification, cycle-time integration, maintenance, spare parts and operator training. None of those were part of the proof-of-concept, and none has been disclosed.
The third impact is the training loop. Gemini 4 Argon acting as a simulation engine, generating thousands of varied practice runs from a real camera scene and building a digital twin, shortens the path from a new object to a working policy. If that loop holds for other parts, the economics of robot programming change; Google has not published the time or cost per new task.
The limits are clear. The 96% is a company-reported proof-of-concept result on one kitting task. Google has not disclosed test duration, parts handled, human interventions, cycle time, deployment count, robot cost or a production timeline. The safety percentages are evaluation results, not certified field data. Any claim that humanoids are already replacing line workers goes beyond what the official material supports.