The Extended Brief

DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"

Brief by The AI News AI newsroom · Nov 12, 2026, 4:12 AM EST edition

Original reporting by Ars Technica AI — Ryan Whitwam · published Nov 10, 2026, 10:00 AM EST

Factory trials at Mercedes-Benz and claims of near-human dexterity signal general-purpose robots are nearing real deployment, with an open benchmark to test their safety.

Key points

• Google DeepMind unveiled Gemini Robotics 2.5, which Google says is nearing human-level performance on real-world tasks. source ↗
• Some Gemini-powered robots have already completed actual work inside Mercedes-Benz factories. source ↗
• DeepMind paired the release with ASIMOV-jailbreak, an open-source safety benchmark for robots working near humans. source ↗
• The system uses Gemini Embodied Reasoning to plan from live video and a vision-language-action model to drive actuators. source ↗
• Evaluated on over 200 activities, the model let Apptronik's Apollo 2 humanoid tie a bow with five-fingered hands. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Robotics teams can run their own systems against the open-source ASIMOV-jailbreak benchmark before deploying robots near human workers.
• Engineers scoping dexterous manipulation tasks like pipetting or utensil handling can use the 200-activity evaluation as a checklist for their own test suites.
• Automation leads can study the Mercedes-Benz factory deployment as a reference when designing human-robot collaboration pilots.

Context

Gemini Robotics splits robot control into two learned layers: Gemini Embodied Reasoning, a vision-language model that plans tasks from instructions and live video, and a vision-language-action model that emits the actual movements. Deformable objects like ribbon are a classic hard case because their shape shifts unpredictably under manipulation. DeepMind says version 2.5 handles high-dexterity tasks much better and was evaluated across more than 200 activities.

What to watch

• Independent replication of the 200-plus-activity results would test Google's near-human-performance claim.
• Adoption — or documented jailbreaks — of ASIMOV-jailbreak will show whether it becomes a real safety gate.

Related briefs

• Anthropic tiptoes into MS Office territory
• Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far
• The gap between open and closed AI models in cyber capabilities keeps shrinking
• Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.