AI News

Google DeepMind Releases VLA Model Gemini Robotics 2.5

The PaperUpdated Nov 11 15:35

AI summary

Google DeepMind released on local time November 10 the new-generation vision-language-action model Gemini Robotics 2.5, which greatly improved the success rate of dexterous manipulation in evaluations of about 200 tasks; for example, using tongs to pick up pastries rose from 5% to 100%, and untying a ribbon bow rose from 0% to 70%. The model completes tasks 20%-30% faster than the previous version, and performance on multi-step and contact-intensive tasks also improved. The team also conducted a proof of concept with Apptronik on Mercedes-Benz production lines in Hungary and Germany, where the humanoid robot Apollo 2 achieved a 96% success rate in picking up and placing automotive metal brackets.

Why read this

SelectedAI score 80

The original text lists the changes in success rates across multiple dexterous manipulation tasks, from which the extent of progress of the VLA model in fine manipulation can be judged.

Topics

Tags

Body · AI translation

Download the app

+1 Google DeepMind releases next-generation VLA model

The Paper reporter Yu Yan 2026-11-11 14:31 Source: The Paper ∙ Future 2% > Listen to full text Font size

On November 10 local time, Google DeepMind officially released its next-generation vision-language-action (VLA) model Gemini Robotics 2.5. In the experimental evaluation data it published, the model's performance on dexterous manipulation is very impressive, and Google says it is "excitingly close to human level."

In evaluations covering about 200 different tasks, the model showed marked improvement on tasks requiring a high degree of dexterity. For example, in the task of using tongs to pick up pastries, the success rate rose from 5% to 100%; ribbons and other deformable objects have always been difficult for robots to manipulate, and the model's success rate in untying ribbon bows rose from 0% to 70%; the success rate for hanging cardigans rose from 0% to 62%; and the success rate for folding pants rose from 0% to 40%. Task precision, robustness and speed all improved significantly.

Compared with previous versions, Gemini Robotics 2.5 completes tasks 20%-30% faster. In addition, Gemini Robotics 2.5 showed clear improvement in multi-step and contact-intensive tasks.

Gemini Robotics 2.5 can untie bows Source: Google DeepMind official website

Google DeepMind revealed that the reason Gemini Robotics 2.5 was able to iterate quickly in a short period of time was largely thanks to the Gemini 4 Argon model released in October 2026. Argon built a "simulation engine" that greatly accelerated the development of Gemini Robotics 2.5.

Google DeepMind robotics research scientist Norman Di Palo further explained on social platforms, "This is our new foundation model, and the first model accelerated through Gemini 4 Argon." He also mentioned, "We are advancing recursive self-improvement (RSI), and in the field of robotics, this is already beginning to show results."

In addition to desktop and everyday household operations, Google has also extended testing to automotive industrial production sites.

Officials revealed that the team, together with the American robotics company Apptronik, carried out proof-of-concept work on Mercedes-Benz automotive production lines in Hungary and Germany.

A humanoid robot equipped with the model completes work on an automotive production line Source: Google DeepMind official website

The Apollo 2 humanoid robot, equipped with the model, takes automotive metal brackets out of bins and then places them into containers in the order required for production, with a success rate of 96% for this task. DeepMind Director Kanishka Rao said directly in an interview with foreign media that closing the final 4% gap will be a "difficult road."

Responsible editor: Huan Yanhong Image editor: Jin Jie Proofreader: Zhang Yan The Paper tip-off: 021-962866 The Paper, unauthorized reproduction is prohibited +1 Favorite I want to report #Google #DeepMind View more

Source

The Paper