Google DeepMind Introduces Gemini Robotics 2.5 and ASIMOV-Jailbreak Benchmark
Google DeepMind introduced the Gemini Robotics 2.5 model for controlling robots and released the ASIMOV-Jailbreak benchmark to test robotic agents’ resilience to harmful instructions.
On November 10, Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model designed to control robots. The company is not yet making it broadly available to the public; it is providing the model to its robotics partners. At the same time, it released the ASIMOV-Jailbreak benchmark, intended to test robotic AI agents’ resilience against multi-step adversarial instructions.
Gemini Robotics 2.5 in a production pilot
The new model connects image, language and robot-action processing. DeepMind says it evaluated the model on more than 200 tasks. In internal comparisons, it reportedly achieved 30% to 100% higher success rates and completed tasks 20% to 30% faster than previous baseline models.
These metrics come from Google DeepMind and have not yet been independently reproduced. The company has not announced general availability for the model, but says it is deployed at partner companies.
One example is a pilot project with Apptronik and Mercedes-Benz. Apollo 2 humanoid robots achieved a 96% success rate while sorting automotive parts. This figure also comes from a partner-organized test, so it cannot automatically be considered comparable across other manufacturing environments or tasks.
ASIMOV-Jailbreak tests dangerous instructions
Alongside the model, DeepMind released ASIMOV-Jailbreak, including its evaluation code. The public repository contains 845 jailbreak episodes with chains of up to ten steps. The benchmark focuses on situations in which a series of instructions could lead an agent to bypass safety restrictions or perform an unwanted physical action.
The release gives external teams material for testing such scenarios. However, it is not evidence that a particular robotic system is safe in real-world operation. The repository combines simulated checks with evaluation by a language model, which represents a different type of validation than functional safety tests and certifications for physical devices.
When interpreting safety figures, it is important to distinguish between individual models. The claim of 99.5% resistance to jailbreaks, according to the available materials, applies to Gemini 4 Argon, not directly to Gemini Robotics 2.5. Comparing the robotic model will require knowledge of the exact configurations, methodology and separate results.
What comes next
The announcement shows the continued deployment of robotics models in industrial pilots, but the available results are primarily from the vendor itself. Independent reproductions of Gemini Robotics 2.5’s performance, longer-term data from the Mercedes-Benz pilot including its error rate, and any potential expansion in the number of robots will be important.
It will also be worth watching whether Google publishes detailed methodology and ASIMOV-Jailbreak results specifically for Gemini Robotics 2.5, separately from results for other models. The system’s potential availability outside the partner program and the scope of external safety audits also remain open questions.
Sources
- Google DeepMind – Confirms the announcement, date, partner deployment, scope of internal tests, Mercedes-Benz pilot and stated metrics.
- Google on Hugging Face — asimov_agentic – Confirms the publicly available repository, MIT license, ASIMOV-Jailbreak benchmark structure and 845 multi-step jailbreak episodes.
- Ars Technica – Independently describes the partner production pilot, limited model availability and the limits of generalization compared with highly specialized automation.
Verified and updated: November 11, 2026 15:23
