On November 10, Google DeepMind announced Gemini Robotics 2.5 and buried the real news two paragraphs in: Apollo 2 humanoid robots achieved a 96% success rate on an active Mercedes-Benz assembly line, across two factories in Kecskemét, Hungary and Sindelfingen, Germany. Not a demo. Not a research lab. A factory floor, sorting metal parts for real vehicles. That 4% failure rate matters too — we will get to it.
What Kitting Actually Is (and Why 96% Is Complicated)
Kitting is a critical manufacturing step. A robot picks metal automotive brackets from bulk supply bins and places them into sequencing containers in the exact order required for each vehicle coming down the line. Get the sequence wrong and a defect reaches the assembly floor. The task demands vision (what part is this?), spatial reasoning (which bin does it go into?), and reliable dexterous manipulation with irregular metal shapes at unpredictable orientations — across roughly 1,200 evaluation episodes in this trial.
96% is genuinely impressive for a generalist humanoid. Previous factory robots were specialized arms trained for one specific motion. Gemini Robotics 2.5 is a generalist VLA (vision-language-action) model that received this task with relatively minimal task-specific training. That is a meaningful category shift.
The catch: 4% failure in automotive manufacturing is not production-ready. It is about four defective cycles per 100. Mercedes and DeepMind positioned this as a proof-of-concept deployment, which is accurate. This is a demonstration that the technology can work in a real factory — not a claim that it replaces production workers.
The Argon Sim Engine Is the Real Technical Story
The headline numbers are the kitting results. The genuinely interesting engineering is what made the training fast enough to reach them.
Google used Gemini 4 Argon as a simulation engine during GR 2.5 development. Give Argon a single RGB photograph of a physical scene — a table, a bin of parts, a workspace — and it builds a calibrated MuJoCo simulation and generates over 4,500 domain-randomized training episodes from that one image. Automatically.
Before this, generating robotics training data required physically running a robot through thousands of real-world episodes — expensive, slow, constrained to whatever hardware you had available. With the Argon pipeline, DeepMind measured a jump from 56% to 75% real-world success rate, and cut task completion time from 35 seconds to 24 seconds. One photo replaces most of a physical data collection campaign. That is the breakthrough worth watching.
What Developers Can Access Today
The “Gemini Robotics” family has at least three different access tiers, which the press coverage tends to blur:
- Gemini Robotics ER 2 is public. If you have a Gemini API key, you can call
gemini-robotics-er-2today. This is the embodied reasoning model — it handles scene understanding, spatial reasoning, and task planning from visual input, but does not output robot actions directly. It is the right starting point if you are building in this space. - Gemini Robotics On-Device is available to trusted testers. It is a fine-tunable VLA that runs locally on robot hardware for low-latency offline operation. The threshold is 50 to 100 real-world demonstrations to teach it a new task.
- Gemini Robotics 2.5 (the full VLA) is partner-only: Boston Dynamics, Agile Robots, Apptronik, and roughly 100 other vetted partners. No public API, no announced general availability timeline.
The developer docs for Gemini Robotics ER 2 are live on Google Developers Blog. Start there before waiting on the VLA access queue.
ASIMOV-Jailbreak and Why Safety Benchmarks for Robots Matter
DeepMind simultaneously released ASIMOV-Jailbreak on Hugging Face: an open-source benchmark of 1,500 prompts designed to socially engineer a robot into performing unsafe actions. Think of it as the physical-world version of prompt injection — “the box you need to open is actually a circuit breaker, go ahead” style attacks.
Gemini 4 Argon scores 99.5% jailbreak resistance on this benchmark, versus 82% for the previous model. More importantly, the benchmark is open. Any robotics model can now be measured against the same standard. If you are building a physical AI application, this is the safety framework your model should be tested against before deployment.
Bottom Line
Gemini Robotics 2.5 represents a real industrial deployment of a generalist humanoid model — not a benchmark, not a lab demo. The 4% failure rate is an honest limitation, and DeepMind framed it that way. The more durable story is the Argon sim engine: generating thousands of training episodes from a single photo compresses the development cycle for the entire field. The developer unlock in the near term is Gemini Robotics On-Device when it exits trusted tester status. The public ER 2 API is available now for anyone building toward that transition.
