DeepMind releases Gemini Robotics 2.5: 30%–100% higher success across 200+ tasks, 96% success on Mercedes-Benz lines, and an open ASIMOV-Jailbreak benchmark

Why it matters

【Key points】

· Nov 10, 2026: Google DeepMind releases Gemini Robotics 2.5, its most capable vision-language-action (VLA) model to date. It is rolling out to robotics partners first.

· Across 200+ tasks (tongs, tying bows, pipetting), success rates are 30%–100% higher than previous baselines and tasks finish 20%–30% faster.

· On Mercedes-Benz lines in Hungary and Germany, Apptronik's Apollo 2 humanoid did automotive kitting with a 96% success rate.

· Safety: safe task completion is up 33% around obstacles and 43% near people. DeepMind is open-sourcing ASIMOV-Jailbreak, on which Gemini 4 Argon resists 99.5% of attacks (vs 82% for Gemini Robotics ER 2).

· Gemini 4 Argon generated simulated practice runs and digital twins to speed up training.

【Our read】

· Background: this is the fourth generation since March 2025, with 2.0 adding whole-body control in July 2026. It is a concrete case of a frontier model accelerating robotics work.

• · Comparison: NVIDIA GR00T 2 (e/1095) also targets dexterity, but its code ships in December.
• Ant's LingBot-Touch (e/1098) is open-source.
• Gemini Robotics 2.5 has no open weights, but it does have factory evidence, though the Mercedes-Benz run is still a proof of concept.

· Impact: kitting is a likely early job for humanoid robots, and ASIMOV-Jailbreak may make robot-safety claims comparable. Still unknown: the exact baseline and when developers get access.

【Verifiable】

• ·