FULL STORY

Google DeepMind Unveils Gemini Robotics 2.5

Google DeepMind released Gemini Robotics 2.5, its latest vision-language-action model, achieving 96% sorting success on Mercedes-Benz lines, followed by a CoRL demo where a humanoid robot tied a bowtie autonomously.

2026-11-10 ~ 2026-11-13 · 2 episodes · 21 posts

Episode 1 · Google DeepMind launches Gemini Robotics 2.5 with 96% pick accuracy on Mercedes lines (2026-11-10, 19 posts)

Google DeepMind released Gemini Robotics 2.5 on November 10, its new vision-language-action model for physical AI, only four months after Gemini Robotics 2. Multiple commentators regard it as the strongest robot model in the series, and it has already moved from lab demos to real Mercedes production lines in Germany and Hungary.

Confirmed

  • The model was tested on 200+ robot tasks covering 200+ complex skills, from tying bows and dispensing snacks with tongs to folding clothes, changing light bulbs, disassembling optical drives, handling deformable and fragile objects, and pipetting; Google frames the gains as breadth, performance and complexity.
  • Per DrSingularity, task success rates rose 30–100% over prior baselines, with 20–30% faster completion.
  • Per Zoubin Ghahramani, the model markedly speeds up skill learning and reduces required training; kimmonismus called the 200+ task demos very impressive.
  • In a POC with Apptronik and Mercedes-Benz, Apollo 2 humanoids pick car brackets from bulk bins and place them in assembly order; keerthanpg notes testing at the Sindelfingen plant, and xmercuryone relays official word that the model also drives Apollo 2 on Hungarian in-production Mercedes lines, sorting brackets per vehicle order at a 96% success rate—still a proof of concept.
  • Per davidpattersonx relaying the official announcement, Gemini Robotics 2.5 is the first robot foundation model "powerfully accelerated" in training by Gemini 4 Argon, with the team explicitly pursuing recursive self-improvement (RSI); 量子位 likewise credits it with RSI capability.

Unconfirmed

  • testingcatalog, bousmalis and a Reddit post (1411337) say the model is built on Gemini 4 Argon; official wording confirms Argon's role in accelerated training and self-improvement, but technical details of "built on Argon" await full documentation.

Why it matters

  • A major upgrade only four months after the previous generation signals sharply accelerating physical-AI iteration.
  • The 96% picking accuracy on real Mercedes lines and 30–100% success-rate gains show such models can perform autonomous work in industrial settings, not just demos.
  • Publicly using Gemini 4 Argon to accelerate robot-model training and pursue RSI marks Google bringing frontier self-improvement into physical AI.

Episode 2 · Google DeepMind Unveils Gemini Robotics 2.5 With Live Demo (2026-11-13, 2 posts)

Google DeepMind launched Gemini Robotics 2.5 at CoRL 2026 with a live demo where a humanoid robot autonomously tied a bow tie and achieved a 96% packing success rate, marking its most capable vision-language-action model yet.