Urgent.News

What's breaking now, across hundreds of outlets.

AI

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

The new model is up to twice as successful at some tasks.

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

Google DeepMind introduced an upgraded version of its AI-powered Gemini Robotics platform, dubbed 2.5. According to the company, this latest iteration of the robotic AI is nearing human-level proficiency in executing real-world tasks. To demonstrate this, some Gemini robots were deployed in Mercedes-Benz factories and successfully completed assigned duties. In the coming days, DeepMind aims to develop an open-source safety benchmark to facilitate the coexistence of robots and humans in shared spaces.

The physical AI platform incorporates multiple computational layers that bridge user input and robotic action. The first layer includes a vision-language model, referred to as Gemini Embodied Reasoning. This component processes instructions using live video feeds to devise a plan for the robot to accomplish the given task. The second crucial component is the vision-language-action (VLA), responsible for generating robotic actions, such as moving an actuator – Gemini Robotics, which has been upgraded to version 2.5.

DeepMind emphasized that Gemini Robotics 2.5 significantly improves at tasks demanding high dexterity. The platform was put through rigorous testing across over 200 activities, ranging from serving food with tongs to using a pipette in a laboratory setting. Google showcased an additional accomplishment, demonstrating how Apptronik’s Apollo 2 humanoid robot, equipped with Gemini Robotics 2.5, can now tie a bow with its five-fingered hands.

This feat is noteworthy as deformable objects like ribbon are traditionally challenging for robots to manipulate.

Rewritten by urgent.news; may contain errors — check the original.

Source: arstechnica.com

Read the original at arstechnica.com →

Report a problem with this story

More in AI

More from Tuesday 10 November →