AI News
All stories
Official

Google DeepMind's Gemini Robotics 2.5 targets robot dexterity, reliability and safety

Google DeepMind describes Gemini Robotics 2.5 as its most capable vision-language-action model, aimed at improving robot dexterity, reliability and safety.

Models · · 3 min read

Google DeepMind has introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action model. The central claim is that robots can learn a wider range of skills faster and then perform them in messier settings than a lab. The company says the model is rolling out to robotics partners. The post itself is the main source for the claims below.

The post covers three areas: dexterity, reliability and safety. On dexterity, DeepMind says it tested the model across 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting a liquid. It reports a 30% to 100% higher success rate than previous baselines, depending on task type, with tasks completed 20% to 30% faster. The post does not name those baselines or say how success was scored. The range is therefore the company's own comparison, not an independent measure.

The reliability claim rests on a single pilot. DeepMind and Apptronik put Apollo 2 humanoids on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. There, the robots performed kitting: picking metal brackets from bulk bins and placing them into sequencing containers in the order each vehicle needed. The division of labor works like this. A separate embodied reasoning model, Gemini Robotics ER, tracks the workflow, while Gemini Robotics 2.5 controls the hands and arms. The post reports a 96% success rate and says the robots were "correcting for any gaps" during picking. Apptronik's chief executive, Jeff Cardenas, said the models improved "dramatically in just a few weeks of post-training" and carried over to Mercedes-Benz production "with very little drop-off." A 96% rate on one kitting task at two sites is a useful data point. But the post does not give the number of attempts, the duration of the trials or the failure modes.

The safety section describes two mechanisms. The first is what DeepMind calls inference-time reasoning guidance, which lets the robot steer around physical hazards while it runs. The company compares it to the safety classifiers that sit around Gemini's chatbot responses. In DeepMind's evaluations, this raised safe task completion by 33% around obstacles and by 43% around nearby people, though the post does not state the baseline for either figure. The second is a jailbreak test. DeepMind is releasing ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. It reports that its Gemini 4 Argon model resists jailbreaks 99.5% of the time, against 82% for the previous baseline, Gemini Robotics ER 2. Both the benchmark and the comparison baseline are DeepMind's own, so these are the company's internal results.

The post also credits Gemini 4 Argon with accelerating the robotics work. According to DeepMind, Argon acts as a simulation engine. From a real camera scene, it can generate thousands of varied simulated practice runs and build a digital twin. Combined with real demonstrations, this helps a robot generalize to new scenes and objects. Argon can also review video of failed tasks to suggest why they failed, then test the best improvements in simulation and directly on real robots. The post does not quantify how much this sped up development, so the acceleration is described but not measured. The announcement closes by framing the release as a step toward what it calls physical AGI, a term the post uses without defining.

What changed

Google DeepMind introduced Gemini Robotics 2.5, described as its most capable vision-language-action model, and it is rolling out to robotics partners.

Type: New capability · Applies to: Gemini Robotics 2.5

Why it matters

Teams building physical AI or robotics products have a new model from Google DeepMind to assess, though the claims so far are general and unquantified.

What you can do

Check the original post for access terms and any published evaluations before planning a robotics pilot.

Original source

Where we saw it

1 source

Google DeepMind

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works