A Guth Labs publication ↗
Back to News

Gemini Robotics 2.5 reaches partners with a 96% factory-pilot result

Conceptual editorial illustration of generic robotic grippers sorting manufactured parts; not an actual factory or product photograph. Branded guthlabs.ai.
AI-generated conceptual editorial illustration | guthlabs.ai

Google DeepMind introduced Gemini Robotics 2.5 on November 10 and began rolling the vision-language-action model out to robotics partners. Its most concrete industrial result is a reported 96% success rate for automotive-parts kitting with Apptronik’s Apollo 2 humanoid at Mercedes-Benz production sites. Google DeepMind announcement

That result makes the release worth watching for manufacturers. It also needs to stay attached to its setting: a proof of concept for a specified workflow, rather than a reliability rating for arbitrary factory work.

A broader manipulation model

Google says it evaluated the model across 200 distinct tasks, including handling tools, delicate objects and flexible materials. In the Mercedes-Benz trial, Gemini Robotics ER tracked the workflow while the 2.5 action model controlled Apollo’s hands and arms. The work took place in Kecskemét, Hungary, and Sindelfingen, Germany. Pilot and evaluation details

The division between planning and motion is important for evaluation. A system can choose the right sequence yet fail at a grasp, or execute a movement correctly while selecting the wrong part. Buyers should ask for those failure categories separately rather than rely on one aggregate score.

Google’s model page describes support ranging from tabletop robots to humanoids and offers an early-access waitlist. The access route is a partner rollout; the public page does not establish unrestricted access to a downloadable model.

Safety testing extends beyond task completion

Google also introduced ASIMOV-Jailbreak and reported improved safety steering around people and obstacles. Its headline 99.5% jailbreak-resistance result applies to Gemini 4 Argon in the agentic evaluation, not a universal safety score for the 2.5 action model. Google’s safety results

The ASIMOV dataset documentation describes paired benign and harmful tasks plus a multi-turn adversarial suite of 845 episodes. It measures both harmful compliance and excessive refusal. The repository carries a CC BY 4.0 label but requires users to accept access conditions and share their Hugging Face information before obtaining the files.

For deployment teams, the next questions concern long runs, human intervention, damaged parts and the time needed to recover from a mistake. A pilot’s success rate is useful evidence. Its operational value depends on the severity and frequency of the remaining failures, and on whether the same performance survives changes in parts, lighting and workflow.

Sources and citations

Revision history

Published .