Google DeepMind Releases Gemini Robotics 2.5: 30% to 100% Higher Success Across 200 Tasks, a 96% Parts-Kitting Rate for Apptronik Apollo 2 on Mercedes-Benz Lines, and an Open-Source ASIMOV-Jailbreak Benchmark

Core Background and Industry Pain Points

The hard part of putting robots into real places is not one polished demo. It is finishing many different tasks reliably in messy settings. Soft or fragile objects, tool use, and safety around people have kept robots in the lab. On October 27, Google Cloud put the embodied-reasoning model Gemini Robotics-ER 2 into enterprise public preview. On November 10, Google DeepMind released Gemini Robotics 2.5, the model that controls the arms and hands.

Architecture Highlights and Underlying Mechanisms

Gemini Robotics 2.5 is a vision-language-action (VLA) model that outputs robot actions directly and works across robot types, including Apptronik's Apollo 2 humanoid (a five-finger hand with 22 degrees of freedom) and the two-arm Franka Duo platform. On the factory line it splits the work with Gemini Robotics ER: ER tracks the workflow and plans steps, while 2.5 controls the hands and arms. For training, DeepMind used Gemini 4 Argon as a simulation engine. Argon rebuilds a digital twin from a single camera view and generates thousands of simulated practice runs, which are combined with real demonstrations. It also watches failure videos, suggests causes, and tests fixes in simulation and on real robots. For safety, the model takes guidance at inference time to steer around obstacles and people.

Authoritative Benchmarks and Measured Comparisons

All figures come from DeepMind. Across more than 200 tasks, the success rate is 30% to 100% higher than earlier baselines, and tasks finish 20% to 30% faster. The parts-kitting proof of concept at two Mercedes-Benz plants reached a 96% success rate. Safety guidance raised safe task completion by 33% around obstacles and 43% near people. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark of hundreds of unsafe requests disguised as normal ones; Gemini 4 Argon resists 99.5% of them, versus 82% for the earlier Gemini Robotics ER 2. The blog does not compare against other companies' VLA models, and there is no third-party replication yet.

Developer Adoption and Getting Started

Gemini Robotics 2.5 is rolling out only to robotics partners, with no public API or weights, so most developers cannot use it yet. Teams that want to prepare can use the Gemini Robotics-ER 2 preview on Google Cloud for task planning and scene understanding, and watch for the open-source ASIMOV-Jailbreak benchmark to check whether their own robot stack can be tricked into unsafe actions. The 96% figure comes from a factory proof of concept, not a production rollout, so retest in your own setting before relying on it.