TL;DR Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026[1]. It's a vision-language-action (VLA) model that expands both the range of skills a robot can learn and how fast it picks them up. In a real-world proof-of-concept, Apptronik's Apollo 2 humanoid used the model to perform automotive kitting on active Mercedes-Benz lines with a 96% success rate[1]. Across 200+ tasks, DeepMind reports a 30% to 100% higher success rate and 20% to 30% faster completion versus previous baselines[1].
A robot representing Google DeepMind's Gemini Robotics 2.5 physical intelligence model

The short answer

Gemini Robotics 2.5 is a step toward "physical AGI" — the idea that robots should be able to handle messy, unstructured, real-world environments rather than only tightly controlled lab settings[1]. DeepMind positions the release around three gains: dexterity (more tasks done better), reliability (dependable work on a live factory floor), and safety (steering around hazards plus a new jailbreak benchmark)[1].

It's rolling out to DeepMind's robotics partners first — not something you can sign up for directly today — but it matters because it's one of the clearest signals yet that frontier model progress is feeding directly into physical, industrial automation[1].

What Gemini Robotics 2.5 is

Like the earlier Gemini Robotics models, 2.5 is a vision-language-action (VLA) model — it takes what a robot sees and turns it into physical actions. It works natively across robot types, including Apptronik's Apollo 2 humanoid (which carries a five-finger, 22 degree-of-freedom SharpaWave hand) and the Franka Duo platform with two-fingered parallel grippers[1].

DeepMind says a large share of the jump came from Gemini 4 Argon, which acted as a simulation engine to accelerate robotics research — generating thousands of varied practice runs from a single real-world camera scene and building realistic digital twins[1].

Dexterity and task breadth

DeepMind evaluated the model across 200 distinct tasks — from serving food with tongs and tying a bow on a gift box to pipetting liquid — which it says represent a step toward human-level dexterity by combining advanced tool use, manipulation of deformable objects, and delicate handling[1]. The headline numbers[1]:

MetricImprovement vs. previous baselines
Task success rate30% to 100% higher (varies by task type)
Task completion speed20% to 30% faster
Tasks evaluated200+ distinct tasks

Source: Google DeepMind's announcement[1].

The Mercedes-Benz proof of concept

The most concrete result is a pilot with Apptronik and Mercedes-Benz on active automotive production lines in Kecskemét, Hungary and Sindelfingen, Germany[1]. Apollo 2 humanoids performed "automotive kitting" — the repetitive, physically taxing job of sorting individual parts from bulk bins into organized trays in the exact order each vehicle needs — reaching a 96% success rate[1].

The system pairs DeepMind's Gemini Robotics ER model to track the workflow with the Gemini Robotics 2.5 actions model to control the hands and arms[1]. DeepMind and Mercedes-Benz both stress that this was a proof of concept, not an announced production deployment[1][2].

Safety: steerability and a new jailbreak benchmark

As robots move near people, physical safety becomes the bottleneck. Gemini Robotics 2.5 adds inference-time reasoning guidance that steers the robot around hazards in real time, boosting safe task completion around obstacles by 33% and around nearby people by 43%[1].

DeepMind also released ASIMOV-Jailbreak, a new open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes. On it, Gemini 4 Argon achieves a 99.5% jailbreak resistance versus 82% for the previous baseline, Gemini Robotics ER 2[1].

Availability

Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners, with Apptronik, Agile, and Mercedes-Benz named as collaborators on the release[1]. There's no public API or consumer access announced as part of this launch[1].

Bottom line

Verdict: Gemini Robotics 2.5 is genuinely notable — not because robots are suddenly flawless, but because a frontier model (Gemini 4 Argon) is now measurably accelerating a downstream product, and the 96% Mercedes-Benz result is a concrete, real-world data point rather than a lab demo. For most businesses the takeaway is directional: physical automation is moving faster than a year ago, and it's worth watching your industry's use cases now[1].

Frequently asked questions

What is Gemini Robotics 2.5?

It's Google DeepMind's latest vision-language-action (VLA) model for robots, released November 10, 2026. It improves dexterity, reliability, and safety for real-world physical tasks[1].

Can I use Gemini Robotics 2.5?

Not directly. It's rolling out to DeepMind's robotics partners such as Apptronik and Mercedes-Benz; no public API or consumer access was announced[1].

What was the Mercedes-Benz result?

Apptronik's Apollo 2 humanoid performed automotive kitting on active production lines with a 96% success rate, using Gemini Robotics ER for workflow tracking and Gemini Robotics 2.5 for manipulation[1].

Sources

  1. Google DeepMind — "Gemini Robotics 2.5 accelerates real-world general dexterity" (Nov 10, 2026) — deepmind.google
  2. Times of AI — "Gemini Robotics 2.5: 96% in Mercedes-Benz Humanoid" (Nov 2026) — timesof.ai