Gemini Robotics 2.5: A 96% Factory Pilot at Mercedes-Benz, and Why the Last 4% Is the Hard Part

A day after Anthropic said its new general-purpose model had more than doubled its robot-control score, Google DeepMind released a model built specifically for robots. On Tuesday, 10 November 2026, the company announced Gemini Robotics 2.5, which it calls its “most capable vision-language-action (VLA) model to date”. The post is signed by Carolina Parada, who leads DeepMind’s robotics work. The headline result comes from outside the lab: Apptronik’s Apollo 2 humanoids ran the model on active Mercedes-Benz production lines and completed an automotive kitting task with a reported 96% success rate.

The figure is impressive. What matters more for anyone thinking about robots in production is what DeepMind’s own robotics leaders said about the remaining 4%.

What was released

Gemini Robotics is split into two layers. A vision-language model, Gemini Robotics ER (embodied reasoning), takes the instruction, watches live video and plans the steps. The VLA, Gemini Robotics, turns that plan into motor commands for arms, hands and grippers. Version 2.5 updates the action layer. It follows Gemini Robotics 2, which DeepMind released in July with whole-body control for humanoids and other robot types.

DeepMind’s claims, confirmed in both its blog and the Ars Technica and The Decoder reports:

  • About 200 evaluation tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid.
  • A 30% to 100% higher success rate than the previous baseline, depending on the task type, and 20% to 30% faster completion. Ars stresses that these are laboratory results.
  • The model is cross-embodiment: it runs on Apollo 2, which has a five-finger, 22-degree-of-freedom SharpaWave hand, and on the two-armed Franka Duo with simple parallel grippers.
  • Availability is limited to partners. No public API or pricing has been announced. DeepMind’s blog thanks Agile, Apptronik and Mercedes-Benz. Ars lists the partners as Agile Robots, Apptronik and Boston Dynamics. The two lists differ, and DeepMind’s own text does not name Boston Dynamics.

The per-task charts in the blog show where the gains come from. Several long-horizon or deformable-object tasks went from near zero to usable. Closing a toolbox lid rose from 0% to 93%, packing pastries with tongs from 5% to 100%, and untying ribbon bows from 0% to 70%. Folding trousers, at 0% to 40%, shows that cloth handling is still far from solved.

The Mercedes-Benz pilot

The pilot ran on production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task was kitting: Apollo 2 picks metal brackets out of bulk bins and places them into sequencing trays in the exact order each vehicle on the line needs. ER tracks the workflow, and the 2.5 VLA controls the hands and arms. DeepMind says the robot corrects gaps as it goes. According to The Decoder, Apptronik CEO Jeff Cardenas said the models improved significantly after “a few weeks of post-training” and moved to Mercedes production with minimal performance loss.

Kitting is a sensible first job for a general-purpose robot. It is repetitive, physically tiring and tolerant of a slower pace, and parts come in many shapes, which is exactly where hard-coded industrial automation is costly to reconfigure.

Training with a frontier model as a simulator

The more interesting part for developers is how the model was built. DeepMind says Gemini 4 Argon, its frontier general model, served as a “simulation engine”. From a single real camera scene it generated thousands of varied practice runs and a digital twin, and these were combined with real teleoperated demonstrations. Argon was also used to diagnose failures from video and propose fixes, which were then tested in simulation and on real robots.

This addresses robotics’ central bottleneck. There is no internet-scale corpus of physical interaction, and teleoperation data is slow and expensive to collect. If frontier models can multiply real demonstrations with credible synthetic variation, the cost of teaching a robot a new task drops. It also mirrors the pattern we covered with Genie 4: world models are becoming training infrastructure, not just demos.

Safety: a robot jailbreak benchmark

DeepMind says Gemini Robotics 2.5 uses inference-time reasoning to steer around hazards, stopping or waiting when something blocks the way. In its own evaluations, this raised safe task completion by 33% around obstacles and 43% around nearby people.

It also released ASIMOV-Jailbreak on Hugging Face, a benchmark of requests that disguise unsafe physical actions. This is a useful contribution, because there is no agreed way to test embodied AI for safety. The published details, however, do not agree with one another:

  • DeepMind’s blog says Argon achieves 99.5% jailbreak resistance against 82% for Gemini Robotics ER 2. Ars reports 98.8% and 73.5%.
  • The blog describes “hundreds of prompts” across “dozens of scenes”. Ars says 1,500 prompts. The Hugging Face dataset card lists a multi-turn suite of 845 episodes plus a few hundred single-turn samples.
  • The dataset is gated behind a Hugging Face sign-in agreement, and its card gives two different licences (CC BY 4.0 in the header, MIT in the README).

None of these discrepancies is damning. They may reflect different benchmark splits or versions. But a safety benchmark is only as useful as its reproducibility, and these should be clarified.

The skeptical read

Three caveats stand out.

First, 96% is not a production number. A human kitting parts gets it right almost every time. DeepMind director Kanishka Rao told Ars that closing the last 4% will be a “steep road”. Parada described the trade-off directly: “We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application.” None of the sources gives the number of trials, the pilot’s duration, how failures were defined or how often a human intervened.

Second, it runs in the cloud. According to Ars, the 2.5 release still runs on cloud servers, so the accuracy and speed gains come with network latency. Ars also notes that the demo video is sped up 2×. For a factory, cloud dependence raises questions about uptime and connectivity on the plant network.

Third, all the figures are self-reported. The lab benchmarks, safety gains and pilot results come from DeepMind and its partners. There is no independent evaluation.

Why it matters

The industry’s bet is that a single general model, trained mostly with synthetic data and adapted in weeks, can replace months of task-specific robot programming. Gemini Robotics 2.5 is the clearest evidence so far that this works at roughly 96% on a real line. It also shows the gap between “works” and “deployable” plainly. For software engineers, the shift is that robot behaviour is becoming a question of post-training, evaluation harnesses and safety benchmarks rather than hand-written motion code. Reliability engineering for probabilistic systems, the discipline the LLM world is still learning, is about to move onto the factory floor.

Sources