Google DeepMind has introduced Gemini Robotics 2.5, its most capable vision-language-action model for robots outside the lab. The company says the release helps machines handle everyday tools and delicate, flexible objects with greater precision in unstructured settings where tasks are less predictable.
Testing centred on dexterity, reliability, and safety across an evaluation set of 200 distinct tasks. The release extends earlier work that brought whole-body intelligence to robots of many types, including humanoids. It also pushes Google deeper into physical AI on the factory floor.
Apollo 2 on live production lines
Two models split the work on Apptronik’s Apollo 2 humanoid during a Mercedes-Benz deployment. The embodied reasoning model Gemini Robotics ER tracks the workflow. The Gemini Robotics 2.5 actions model controls the robot’s hands and arms.
At plants in Kecskemét, Hungary, and Sindelfingen, Germany, Apollo 2 performed automotive kitting on active production lines, sorting parts from bulk bins into organised trays in the exact order each vehicle required. This wasn’t a laboratory demo.
Apollo 2 picked and packed automotive parts with a 96 percent success rate during the proof-of-concept. Google DeepMind says robots could take on the monotonous work so technicians can focus on complex build tasks.
Apptronik said in a statement: “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.”
Benchmark performance across hardware types
Gemini Robotics 2.5 achieves a 30 percent to 100 percent higher success rate than previous baselines, depending on task type, completing the same evaluations 20-30 percent faster. Google reports improved task precision alongside those gains. The company says the model increases both the breadth of skills a robot can learn and the speed at which it acquires them.
The model family also spans hardware beyond humanoids. Apollo 2’s SharpaWave hand offers 22 degrees of freedom; the same software also runs on the Franka Duo platform, which uses parallel grippers. Gemini Robotics 2.5 doesn’t need a humanoid hand, so it isn’t tied to one robot design.
Simulation training and jailbreak safety
Google credits Gemini 4 Argon with accelerating the underlying research. The system worked as a simulation engine, generating practice runs from camera scenes and building a digital twin of each task so engineers could test why a task failed before applying changes on real robots. That simulation-first method reflects broader enterprise interest in world models.
Inference-time reasoning guidance now steers robots around physical hazards as they move into everyday environments. The new ASIMOV-Jailbreak benchmark measures resistance to disguised unsafe requests across varied scenes.
Argon achieved a 99.5 percent jailbreak resistance rate on the benchmark, a figure Google describes as a substantial improvement over the previous Gemini Robotics ER baseline. It was the strongest resistance result ever recorded.
See also: Semtech SGP.32 delivers one global eSIM for IoT fleets
Want to learn more about the IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including AI & Big Data Expo and the Cyber Security Expo. Click here for more information.
IoT News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
