Loose metal brackets in a bulk bin still beat most robots. Parts shift, trays change, people walk by.
Google DeepMind says its new Gemini Robotics 2.5 closes that gap. A vision-language-action model is a system that turns camera views and language instructions directly into arm and hand motions, and DeepMind tested this version on more than 200 distinct tasks, from serving food with tongs to tying a bow to pipetting liquid.
Speed is the pitch.
The robot meets a job in two layers. Gemini Robotics ER tracks the multi-step workflow, like what part goes in what tray slot next, while Gemini Robotics 2.5 drives the hands and arms to grasp and place. To teach those motions faster, Gemini 4 Argon acts as a simulation engine that builds a realistic digital twin from a real camera scene and then generates thousands of varied practice runs, which DeepMind says helps the policy generalize to new objects with far less physical data and lets researchers replay failures in video and test fixes in simulation before touching hardware.
That combination is aimed at dexterity with ordinary tools and soft materials. DeepMind reports 30% to 100% higher success than prior baselines depending on task type, according to DeepMind, while finishing 20% to 30% faster, according to DeepMind. The same model runs very different bodies, including the two-fingered Franka Duo grippers and Apptronik Apollo 2 with the five-finger, 22 degree-of-freedom SharpaWave hand, according to DeepMind.
The factory test was deliberately boring.
On active Mercedes-Benz lines in Kecskemet, Hungary and Sindelfingen, Germany, Apollo 2 did automotive kitting on its own, reaching into supply bins and sorting brackets into sequencing containers in vehicle order, and hit 96% success, according to DeepMind. Jeff Cardenas, Apptronik co-founder and CEO whose humanoids did the work, said the models improved dramatically in weeks of post-training and carried over to Mercedes-Benz production with very little drop-off. That is still a proof-of-concept for repetitive prep work, not a full shift replacement.
Safety got its own upgrade because homes and factories are shared spaces. Inference-time reasoning guidance is live steering that nudges motion around physical hazards, and DeepMind says it lifted safe completion by 33% around obstacles and 43% around nearby people, according to DeepMind. For language tricks, DeepMind is releasing ASIMOV-Jailbreak, an open benchmark with hundreds of disguised unsafe requests across dozens of scenes, where Gemini 4 Argon refused 99.5% of attacks versus 82% for the prior Gemini Robotics ER 2 baseline, according to DeepMind. The timing lines up with a broader push to score jailbreaks like vulnerabilities, as the Frontier Model Forum has proposed a CVSS-style rating framework for capability gain and deployability backed since 2023 by Anthropic, Google, Microsoft and OpenAI, per Tom's Hardware.
Watch for rollout to robotics partners and whether that Argon simulation loop, earlier used to take heavy 3D work off local hardware, keeps cutting the data cost of each new grip.
