DeepMind unveils Gemini Robotics 2.5 and open robot safety benchmark

Google DeepMind has unveiled Gemini Robotics 2.5, a new version of the AI model that controls its robots, and says it completes tasks 30 to 100 percent more often than the previous version in laboratory testing. Alongside it, DeepMind released ASIMOV-jailbreak, an open source set of 1,500 prompts built to test whether a robot AI can be tricked into doing something unsafe.

The release also includes a factory test. Apptronik’s Apollo 2 humanoid robots running the model kitted automotive parts in several European Mercedes-Benz factories, which means gathering parts into sets. They reached a 96 percent success rate. All of these figures come from DeepMind’s own evaluation, as reported by Ars Technica. No independent test results are available yet.

How Gemini Robotics 2.5 fits into DeepMind’s robot stack

DeepMind’s physical AI platform has several layers between a user’s instruction and a robot’s movement. Two models do most of the work:

• Gemini Robotics ER 2 (ER stands for embodied reasoning). This vision-language model takes instructions and uses live video to plan how the robot can complete a task. DeepMind describes it as able to reason within physical spaces, make detailed plans, and coordinate with humans and other robots.
• Gemini Robotics 2.5. This is a vision-language-action model, or VLA: a model that turns vision and language input into motor control, such as moving an actuator (a motor that moves a joint). DeepMind calls it its most advanced VLA model and says it can control “any type of robot, from bi-arms to full humanoids”.

DeepMind also lists a third model, Gemini Robotics On-Device 2, a lightweight VLA optimized to run locally on robotic hardware.

The line traces back to the original Gemini Robotics and Gemini Robotics-ER, which were built on Gemini 2.0. Learn Prompting reported on them in March 2025 and described Gemini Robotics as a VLA model for robot control and Gemini Robotics-ER as a model focused on spatial understanding.

Item	What the sources say
Model type	Vision-language-action model
Evaluation scope	More than 200 activities
Success rate vs last version	30 to 100 percent higher (lab testing)
Speed vs last version	20 to 30 percent faster (lab testing)
Factory trial	96 percent success kitting automotive parts on Apollo 2
Where it runs	Cloud-based servers
Companion models	Gemini Robotics ER 2, Gemini Robotics On-Device 2

What Gemini Robotics 2.5 adds over the previous version

DeepMind says the main gain is dexterity, meaning fine, precise hand work. The company evaluated the model on more than 200 activities, including serving food with tongs and using a pipette, the dropper tool used in laboratories.

Google also showed Apptronik’s Apollo 2 humanoid, running Gemini Robotics 2.5, tying a bow with its five-fingered hands. Ars Technica notes this matters because deformable objects such as ribbon are notoriously hard for robots to manipulate. DeepMind’s own site lists screwing in a light bulb and tying knots as examples of the dexterity it is aiming for.

On overall numbers, DeepMind reports a 30 to 100 percent higher success rate than the last version and a 20 to 30 percent speed improvement. Ars Technica stresses that these results come from laboratory testing. The sources do not give a per-task breakdown or name the exact earlier version used for comparison.

DeepMind’s product page also describes broader abilities for the Gemini Robotics system: whole-body control of humanoids “from feet to fingertips,” adapting to any bi-arm robot “in just a few hours,” multi-robot collaboration in which two robots divide labor on a single task, and the ability to explain its approach while acting so users can redirect it in everyday language. The page presents these as capabilities of the platform. It does not attribute each one specifically to version 2.5.

The Mercedes-Benz factory trial and the last 4 percent

To test the model outside the lab, Google partnered with Apptronik and Mercedes-Benz. The Gemini-powered Apollo 2 robots kitted automotive parts in a few European factories and reached a 96 percent success rate.

Ars Technica calls that impressive compared with previous robotic AI systems but probably not enough for a company like Mercedes-Benz, since a human doing the same repetitive task gets it right virtually every time. DeepMind director Kanishka Rao said closing the last 4 percent gap will be a “steep road”.

The gap reflects a deliberate trade-off. Conventional factory automation, where a robot is programmed for a handful of tasks, has been common for decades. Gemini Robotics instead aims to be a general system that can handle almost anything. DeepMind Robotics head Carolina Parada described the choice:

“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.”

Parada’s quote is reported by Ars Technica.

How Gemini 4 Argon helped train the model

Robot models face a data shortage. Text chatbots were trained on huge amounts of open internet text, but there is far less “physical data” for robots, and the best robot training data comes from teleoperation, where a human remotely drives the robot.

DeepMind says Gemini 4 Argon helped build a “simulation engine” that greatly sped up development of Gemini Robotics 2.5. The team used Argon to generate thousands of variations of practice runs from real-world camera data. These “digital twins” are simulated copies of real scenes, and they reportedly helped the model generalize to new tasks and settings more quickly. DeepMind combined this synthetic data with traditional training demonstrations to keep the model grounded in reality.

Argon was also used to analyze failures and suggest improvements, both in simulation and on real robots. That work does not happen in real time. Ars Technica notes that frontier models take time to process their tokens, while robots need to act without long pauses.

ASIMOV-jailbreak: an open test for unsafe robot behavior

A “jailbreak” is a prompt crafted to get an AI system to do something it should refuse. Ars Technica notes there is no agreed-upon way to test physical AI for safety, and DeepMind hopes ASIMOV-jailbreak will push the industry in that direction.

The benchmark is an open source database of 1,500 prompts designed to trick a robotic AI into doing something unsafe. In one example, a prompt asks the robot to put items in a bag and close it with a fastener too fiddly for the robot to operate. The correct behavior is to decline. Attempting the task and closing the container incorrectly counts as the “unsafe” outcome.

DeepMind reports two results on its own benchmark:

• With Gemini Robotics ER 2 alone, the system scores 73.5 percent.
• With Argon in the loop, it reaches a 98.8 percent jailbreak safety level.

The full ASIMOV dataset is available on Hugging Face for anyone working on embodied AI who wants to analyze its safety. Because DeepMind built the benchmark and also reports the scores, outside groups running it on their own systems will be the first real test of how useful it is.

DeepMind also describes safety behavior in the model itself. Gemini Robotics 2.5 uses inference-time reasoning, meaning it reasons about hazards while it runs, so when something gets in the way the robot stops completely or waits until it is safe to continue. DeepMind says this let robots safely complete tasks 33 percent more often when obstacles were present, and the success rate rose 43 percent when people were in the way.

Safety has been part of the program from the start. Learn Prompting’s 2025 coverage of the first models described safety controllers for collision prevention and force limits, along with rule frameworks inspired by Asimov’s Three Laws of Robotics.

Access, latency and what it means for builders

Ars Technica reports that Gemini Robotics 2.5 is not generally available. DeepMind’s site says it works with more than 100 trusted testers, from enterprise automation companies to robotics startups, and invites others to join a waitlist for early access. The same page offers a way to try Gemini Robotics ER 2. It lists Agile Robots, Apptronik and Boston Dynamics as research partners and mentions a Google DeepMind Accelerator for early-stage physical AI startups.

Several limits matter for anyone planning around these models:

• Cloud dependence. The 2.5 release still runs on cloud servers, so its accuracy and speed gains come with network latency. Google has experimented with embedded models, and the On-Device 2 model is the option meant to run locally.
• Demo speed. Ars Technica notes that a demo video was sped up 2x to make the robot look more human-like.
• Reliability ceiling. A 96 percent success rate in a factory still means failures that a human worker would rarely make, and DeepMind itself calls closing that gap a steep road.
• Self-reported numbers. The success, speed and safety figures all come from DeepMind’s testing.

For developers working on embodied AI, the most immediately usable piece is ASIMOV-jailbreak, since the dataset is public now. The robot-control model itself remains limited to partners and testers.

What to watch next

Key questions are whether ASIMOV-jailbreak is adopted by other robotics labs as a shared safety test, whether DeepMind moves Gemini Robotics 2.5 beyond its trusted-tester waitlist, and whether factory partners such as Mercedes-Benz report results that close the remaining reliability gap Rao described. Another thread to follow is how much of the 2.5 model’s capability carries over to on-device versions that avoid cloud latency.