Google DeepMind's latest robot model learns hundreds of tasks, from folding laundry to sorting car parts on a factory line
Google DeepMind has unveiled Gemini Robotics 2.5, a new model built to teach robots more skills, faster. The company is calling it its most capable vision-language-action model so far.
The pitch is broad competence over party tricks. Demonstrations covered hundreds of everyday and industrial jobs, from folding laundry to sorting automotive parts.
What Gemini Robotics 2.5 actually does #
Google DeepMind announced the model on November 10, 2026. It is a vision-language-action model, or VLA, which is the category the company is pushing hard.
The headline goal of this release is twofold. Google DeepMind wants robots to learn a wider range of tasks, and it wants them to pick those tasks up more quickly.
To test that, the company evaluated the model on 200 distinct tasks. Those tests emphasized advanced tool use and delicate handling, the kind of work where a clumsy grip ruins the outcome.
The performance numbers are the part that stands out. Compared with earlier models, success rates improved by 30% to 100% across various tasks.
Speed moved too. Task completion ran 20% to 30% faster than prior iterations.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The Mercedes-Benz test run #
The most concrete result came from a factory floor rather than a lab. Apptronik’s Apollo 2 humanoid robot ran a proof-of-concept on a Mercedes-Benz production line.
Apollo 2 handled sorting and packing car components. It hit a 96% success rate during that trial.
The model is rolling out to partners in the robotics industry. The company is presenting Gemini Robotics 2.5 as a versatile dexterity layer that can work with different robotic architectures.
Safety gets its own upgrade #
Capability was only half the announcement. Google DeepMind also highlighted improvements to safety in robotic operations.
The model features refined steerability, meaning operators have better control over how the robot behaves. It also brings stronger resistance to adversarial manipulation, the robotics version of someone trying to trick an AI into misbehaving.
The release includes a new benchmark for adversarial defenses in robotics.
How the Gemini Robotics line got here #
Gemini Robotics 2.5 builds on the Gemini Robotics 2 family, which Google DeepMind introduced on July 30, 2026. That earlier release included whole-body control for humanoid robots.
The 2.5 release layers better dexterity and faster learning on top of that foundation.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.