By RWT Robotics Editorial Team Published
DEMO ONLY
The only account of the pilot is DeepMind’s own, from a single proof-of-concept, and none of the sources we found give a success rate, a trial count or any data on human interventions. That makes this a strong demo, not a deployment.
What we know
Google DeepMind published a post announcing Gemini Robotics 2. The company confirmed it is rolling the models out to robotics partners. This is not a public release, and nothing in the post as we have it suggests one is coming soon.
The pilot account comes from that same post. DeepMind says its models helped Apptronik’s Apollo 2 humanoid pick and pack automotive parts during a proof-of-concept on Mercedes-Benz production lines. DeepMind describes the kitting as autonomous. That is DeepMind’s characterization, and nobody outside the company has verified the result.
The task itself is specific, which helps. According to DeepMind’s blog, as reported by The Decoder, the pilot ran at Mercedes-Benz sites in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 picked metal brackets from supply bins and placed them into sequencing containers in vehicle order.
So yes, to answer the obvious question: by DeepMind’s account, this happened in real Mercedes-Benz plants, not in a Google lab mockup. But “on production lines” and “in production” are different claims. DeepMind calls it a proof-of-concept, and we should take that label at face value.
The partnership is not new, either. Apptronik and Mercedes-Benz have worked together on Apollo humanoids before, so the plants were not a cold start for the robot vendor.
What we don’t know yet
The biggest gap is performance data. The sources we found give no success rate or trial count for the pilot. Without those, there is no way to judge how reliable the robot actually was on the line.
We also have no data on human interventions. “Autonomous” carries a lot of weight in the DeepMind post, and without intervention counts there is no way to check what it covers. Did an operator reset bins? Clear jams? Restart a stalled run? We don’t know.
Several more numbers are circulating, and we could only reach them through secondary summaries of DeepMind’s post. That coverage says the models were tested on more than 200 tasks, with success rates improved by 30% to 100% over earlier models. It also says safe completion around obstacles improved 33% and around people 43%. We have not confirmed these figures against the original. They also describe a broader benchmark, not the Mercedes pilot.
Secondary coverage also says the work was split between two models. Gemini Robotics-ER reportedly tracked the workflow, and Gemini Robotics 2 controlled the hands and arms. That architecture is plausible and fits how DeepMind has described its stack before, but we are relaying it secondhand.
Then there is the model lineup. We could not cleanly resolve which of DeepMind’s models ran in the Mercedes pilot. We have left that question open rather than guess.
What is rumored
The Decoder reports that DeepMind used a simulation engine called “Gemini 4 Argon” during development. The Decoder is a generally reliable AI news outlet, but this is a single-source detail, and we found no corroboration anywhere else. We grade it D. Until DeepMind names the tool itself, treat “Argon” as an unconfirmed claim about the training pipeline, not a known product.
What it means
Start with what is genuinely encouraging. Kitting brackets into sequencing containers is a real factory job. It feeds the line in vehicle order, so the robot has to get the right part into the right slot, not just pick something up. Running it at two working plants is a better test than a curated tabletop demo, and choosing a narrow, well-defined task is the right way to run a first industrial pilot.
Now the skepticism. In sequencing, a wrong or missing bracket does not stay local. It shows up downstream at the station that needed it. Whether the robot’s performance is acceptable depends on how often it failed, what a failure looked like, how it was caught, and who fixed it. None of that is in the reporting we have.
On the angle we set out to check: the pilot did run at real Mercedes-Benz sites, per DeepMind. But it has not been tested outside Google’s own reporting. There is no independent replication, no published success rate or trial count, no intervention data, and access is limited to partners. Outside researchers cannot run the models and check them. For now, “how does it hold up outside Google’s setup” has no answer, because nobody outside Google and its partners has been able to try.
What would change our read: DeepMind or Mercedes-Benz publishing success rates, trial counts and intervention rates for the pilot, a partner reporting results on its own hardware and tasks, or Mercedes-Benz describing a move from proof-of-concept to sustained use on the line. Any one of those would turn this from a promising demo into evidence.
Until then, file it where DeepMind itself filed it: a proof-of-concept with a promising story attached, reported by the company that built the model.
Keep reading
Robotics NewsRing Robotics debuts with a $249.99 indoor drone and a $999.99 patrol robotRing Robotics launched a $249.99 indoor Drone Cam awaiting FCC approval and a $999.99 Rover Cam patrol robot,…Nov 10, 2026 · 5 min read
Robotics NewsHumanoid robot Qestr quits Antarctic Ice Ultra after about 3 km, Reuters reportsQestr, the first humanoid in the Antarctic Ice Ultra, quit after about 3 km. Autonomy, builder and cause…Nov 10, 2026 · 4 min read
Robotics NewsNvidia says GR00T 2 learns from human video, but gives no release dateNvidia says GR00T 2 hits 71% on one demo per task. The figures are company claims, and no…Nov 9, 2026 · 4 min readCredits and sources
By RWT Robotics Editorial Team
- Published
