A 96% sorting success rate—and a bedsheet that still isn’t folded

On November 10, Google DeepMind released Gemini Robotics 2.5, calling it the strongest vision-language-action model (VLA) to date. It can look at images, understand instructions, and then turn those instructions into robot actions. The team evaluated it on more than 200 different tasks; task success rates were 30% to 100% higher than the previous-generation baseline, and completion was 20% to 30% faster.

The demo tasks include folding clothes, serving food with tongs, tying bows, pipetting, and sorting car parts on a factory floor. The last one is easier to underestimate. Google DeepMind ran a proof of concept with Apptronik and Mercedes-Benz: an Apollo 2 humanoid robot used Gemini Robotics 2.5 for automotive parts kitting, placing loose parts into trays in assembly order, with a 96% success rate. Test locations included Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany.

On the hardware side, Gemini Robotics 2.5 natively supports multiple robots, including the Apptronik Apollo 2 humanoid robot, equipped with 22-DOF SharpaWave hands, and the Franka Duo dual-arm platform. The official emphasis is on three areas of progress: dexterity, reliability, and safety. Dexterity is about whether it can use tools, handle deformable objects and fine objects; reliability is about continuous work in a factory; safety is about whether, once the robot enters real spaces, it can identify hazards, avoid people, and resist jailbreak-style instructions.

The training approach is worth a look for readers with a development background. Gemini 4 Argon is used as a simulation engine: it generates a large number of simulated practice spaces from a single real camera viewpoint, builds a digital twin, and then combines that with real demonstrations to help the robot adapt to new scenes and new objects. It can also watch failure videos, infer the cause of failure, and test improvements in simulation and on the real robot. Google DeepMind says this helped Gemini Robotics 2.5 shorten training time that would otherwise have taken months.

The safety section also has concrete numbers. Gemini Robotics 2.5 uses inference-time guidance to avoid physical hazards in real time. In official evaluations, completion rate on safety tasks involving going around obstacles improved by 33%, and in scenarios with people nearby it improved by 43%. They also released the ASIMOV-Jailbreak benchmark, containing hundreds of unsafe instructions disguised as normal requests and covering dozens of scenarios. Gemini 4 Argon's jailbreak resistance rate on this benchmark is 99.5%, compared with 82% for the previous-generation Gemini Robotics ER 2.

The success-rate charts are out too. On multi-step, contact-heavy tasks, closing a black toolbox lid went from 0% to 93%, taking a cup off a hook and placing it on a table from 29% to 75%, and sweeping trash into a dustpan with a brush from 0% to 52%. In dexterous manipulation tasks closer to daily life, using tongs to serve pastries went from 5% to 100%, untying a ribbon bow from 0% to 70%, hanging up a cardigan from 0% to 62%, and folding pants from 0% to 40%.

Taken together, these numbers are not all worth the same. Factory sorting is a structured environment with repetitive tasks and verifiable results; folding clothes involves deformable objects and unstructured environments, and demo success does not mean it is usable for housework. In the comments, some asked whether it can handle clothes that are already in a messy pile, some asked whether folding includes putting them away in the closet, and others said the real test is a fitted sheet. These voices are closer to the home setting than the cuts in the launch video.

Alan D. Thompson of LifeArchitect.ai marked this release as a milestone in his AGI countdown. His assessment: there are clear gains in dexterity and factory reliability, but multi-step fastener assembly—like IKEA furniture—has not appeared yet, so that last bit remains open. This perspective is more interesting than just looking at 'it can fold clothes.' Robots enter factories first, take over repetitive and physically demanding work, and then gradually enter homes; that is roughly the path. For developers, what is worth watching is how VLA, simulation closed loops, and safe steerability stack together; for ordinary users, the folding-clothes videos are worth watching, but ordering one is still a long way off.

Official blog: https://goo.gle/4zBBUA6