Google DeepMind Releases Gemini Robotics 2.5 and Opens Safety Benchmark for Embodied AI
Google DeepMind published Gemini Robotics 2.5 on Tuesday (November 10), which the company describes as its most capable vision-language-action (VLA) model for physical robot control to date, along with the open-source ASIMOV-Jailbreak benchmark designed to measure the robustness of embodied AI agents against adversarial prompts.
What Gemini Robotics 2.5 brings
Unlike its predecessor Gemini Robotics 2 (released in July 2026), version 2.5 was designed to handle deformable objects, tools, and fragile items — widely considered the biggest barriers to deploying humanoid robots in unstructured industrial environments. In DeepMind's evaluations, the model outperformed existing baselines by 30% to 100% depending on task type, and completed tasks 20% to 30% faster.
The numbers are striking for tasks that Gemini Robotics 2 could not perform at all: closing a black toolbox lid (0% → 93%), packing pastries with tongs (5% → 100%), untying ribbon bows (0% → 70%), hanging a cardigan (0% → 62%). Even on the most complex tasks, the 2.5 achieves success rates from 40% to 75%, which still leaves room for improvement but signals a significant leap forward.
The model was evaluated on over 200 distinct tasks, including serving food with tongs, tying bows on gift boxes, and performing liquid pipetting — tasks that require fine coordination between vision, planning, and motor action.
Split architecture: ER + 2.5
A key architectural innovation is the separation between two models: Gemini Robotics ER (Embodied Reasoning), which handles planning and workflow tracking, and Gemini Robotics 2.5, which controls the robot's hands and arms. This division allows the system to track long-horizon tasks while executing complex actions in real time.
DeepMind tested the architecture with the Apptronik Apollo 2 humanoid robot (five-finger hand with 22 degrees of freedom from SharpaWave) and the Franka Duo platform (two-finger parallel grippers). The model also works natively with more than one robot body type.
On the factory floor: Mercedes-Benz and Apptronik
The most concrete test took place on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, where Apollo 2 performed automotive kitting: picking automotive parts from bulk bins and organizing them into sequential trays in the exact order needed for each vehicle on the line. The success rate was 96%.
> "This is what we built Apollo for: useful work alongside people in real industrial settings. Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off. This is the foundation for bringing humanoid robots to industrial and commercial environments at scale" — Jeff Cardenas, CEO of Apptronik.
The test is described as a proof of concept, and DeepMind has not yet disclosed plans for large-scale deployment or commercial release dates.
Safety as a priority
With robots operating in unpredictable environments, DeepMind has focused heavily on safety. Gemini Robotics 2.5 uses inference-time reasoning to steer around physical hazards in real time — a guidance system that works similarly to safety classifiers added to LLM responses, but applied to physical motor actions.
In evaluations, the guidance increased safe task completion by 33% around obstacles and 43% around people. The Gemini 4 Argon model leads Google's agentic robotics safety evaluations with a 99.5% jailbreak resistance rate — a significant improvement over the 82% rate of Gemini Robotics ER 2.
The ASIMOV-Jailbreak benchmark has been released on Hugging Face with hundreds of prompts that disguise unsafe requests across dozens of different scenarios, allowing the community to test the robustness of their own embodied models.
Gemini 4 Argon as a simulation engine
A less-publicized but potentially more impactful aspect: Gemini 4 Argon, the model focused on robotics research, is being used as a simulation engine. From real-world camera scenes, Argon builds a digital twin and generates thousands of varied simulated practice runs. Combined with real demonstrations, this enables the robot to generalize to new scenes and objects much faster.
Argon also helps researchers hypothesize why failures occur from video recordings and test improvements both in simulation and directly on real robots.
What to expect next
Gemini Robotics 2.5 is not available as a public API — access is through robotics partners, including Apptronik, Mercedes-Benz, and Agile Robotics. There is no general release date, pricing, or public timeline.
The trend is clear: the line between "digital" AI and "physical" robotics is dissolving. Google is positioning Gemini 4 Argon as the research model that accelerates the entire development cycle, from simulation to deployment on real robots. And the open ASIMOV-Jailbreak benchmark signals an intent to build an ecosystem — the more teams test model safety, the more reliable these systems will become when they leave the lab.
The question is whether large-scale industrial adoption of humanoid robots will depend more on advances in motor dexterity or on gains in safety and reliability — and whether the market will mature before the underlying technologies catch up.
---
Sources: DeepMind Blog (deepmind.google), CellCog (cellcog.ai), Coding is Love (codingislove.com)
Sources: DeepMind Blog, CellCog, Coding is Love
✓ Independent sources cross-checked and verified before publishing
