The Information Machine
ConcludedRan 1 dayin the edition since 11 Nov 20264 sourcesupdated 11 Nov 2026

Google DeepMind's Gemini Robotics 2.5 release

The gist

Latest: Google DeepMind Releases Gemini Robotics 2.5 Vision-Language-Action Model

The Mercedes-Benz deployment provides a concrete industrial test of whether general-purpose robot models can operate reliably in production environments. The release of ASIMOV-Jailbreak as an open benchmark gives the broader field a shared tool for evaluating robot safety against adversarial prompting.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model for physical AI. The model achieves 30-100% higher task success rates and completes tasks 20-30% faster than previous baselines across more than 200 distinct tasks. The architecture separates reasoning, handled by Gemini Embodied Reasoning, from action generation, handled by Gemini Robotics, in distinct computational layers.

The most prominent deployment is an industrial proof-of-concept at Mercedes-Benz production lines in Hungary and Germany, where Apptronik's Apollo 2 humanoid robots running the model achieved a 96% success rate on automotive kitting tasks. Demonstrated capabilities include folding laundry, serving pastries, sorting car parts on a factory floor, and tying a bow with five-fingered hands. DeepMind noted that deformable objects like ribbon are notoriously difficult for robots to manipulate.

On safety, new inference-time safety reasoning boosted safe task completion around obstacles by 33% and around nearby people by 43%. DeepMind released ASIMOV-Jailbreak as an open-source benchmark for jailbreak robustness testing; Gemini 4 Argon scored 99.5% resistance on it compared to 82% for the prior baseline Gemini Robotics ER 2. Gemini 4 Argon also served as a simulation engine during development, generating thousands of varied simulated practice runs from real-world camera scenes to build digital twins.

Embodied reasoning features from Gemini 2.5 Flash and Pro became available through Google AI Studio, the Gemini API, and Vertex AI as of June 2025, with a Gemini Robotics-ER model accessible via a trusted tester program.

How it developed
First in the edition · 11 November 2026
10 November 2026in the 11 Nov edition

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model that separates reasoning from action generation in distinct computational layers, reporting 30-100% higher task success rates and 20-30% faster completion across more than 200 tasks compared to previous baselines.

The most prominent deployment is a proof-of-concept at Mercedes-Benz production lines in Hungary and Germany, where Apptronik's Apollo 2 humanoid robots achieved a 96% success rate on automotive kitting tasks. New inference-time safety reasoning boosted safe task completion around obstacles by 33% and around nearby people by 43%, and DeepMind released ASIMOV-Jailbreak as an open-source jailbreak robustness benchmark. Embodied reasoning features from Gemini 2.5 Flash and Pro had been available through Google AI Studio and the Gemini API since June 2025, extending the model family's timeline back over a year before this release.

24 June 2025

Gemini 2.5 Flash and Pro embodied reasoning features made available via Google AI Studio, Gemini API, and Vertex AI; Gemini Robotics-ER accessible via trusted tester program.

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy