Robotics Gemini: New Google DeepMind for robots

The models are both designed on Google Gemini, which is the multimedia basic model that can process text, sound and images to answer questions and provide advice and assistance in general. Deepmind calls for the first new models, Gemini Robotics, a “Language-Action”, which means that it can take all these inputs itself and then take out the material procedures instructions for the robot. Models are designed to work with any device system, but they are often tested on gunmen Aloe 2 The system presented by Deepmind last year.

In a demonstration video, Voice says: “Pick a basketball and rose” (at 2:27 in the video below). Then the robot arm takes carefully a miniature basketball and drops it into a mini network-and although it was not Donk at the American Professional League, it was enough to stir deep researchers.

https://www.youtube.com/watch?Google DeepMind has released this experimental video that displays the capabilities of its Gemini Robotics model to control robots.Gemini robots

“For example, this basketball is one of my favorites,” he said Kanishka RaoThe main software engineer, at a press conference. He explains that the robot has never seen anything related to basketball, “but the basic basis model that had a general understanding of the game, knows how the basketball network looks, and understood what the term” Slam Dunk “means. So the robot was able to deliver them [concepts] “To accomplish the task in the material world,” says Rao.

What is the progress of Gemini robots?

Carolina ParadaGoogle DeepMind said in a briefing that new models improve the previous robots of the company in three dimensions: generalization, ability to adapt, and ingenuity. She said that all these developments are necessary to create a “new generation of useful robots.”

Circular means that the robot can Apply the concept of his learning in one context to another position, and the researchers looked at the optical circular (for example, is it confused if the color of an object or background changes), and the generalization of instructions (can it explain the orders that are formulated in different ways), and the circular of the procedure (can a procedure not have previously been done before).

Parada also says that the robots with Gemini can adapt to the instructions and changing conditions. To prove this point in a video clip, one of the researchers told a robot to put a set of plastic grapes in the clear Tupperware containing, then moved to changing three containers on the table Shell game. The robot arm follows the clear container around it so that it can direct it.

https://www.youtube.com/watch?GOMINI ROBITICS says that GIMINI ROBITICS is better than previous models in adapting to the instructions and changing conditions.Google DeepMind

As for ingenuity, experimental videos showed automatic weapons folding a piece of paper in an origami fox and performing other accurate tasks. However, it is important to note that the impressive performance here is in the context A narrow range of high -quality data that has been trained in these specified tasks, so the level of ingenuity represented by these tasks is not generalized.

What is the embodied logic?

The second model presented today is Robotics Gemini, with ER of “embodied thinking”, a type of intuitive material world that understands that humans are developing with experience over time. We are able to do smart things like taking a look at an object that we have never seen before and we are guessing educated about the best way to interact with it, and this is what DeepMind seeks to simulate the Gemini Robotics-a.

Prada gave an example of the ability of robots-Air to determine an appropriate absorption point for capturing a cup of coffee. The handle is properly determined, because this is the place where humans tend to understand the cups of coffee. However, this shows a possible weakness in relying on human training data: for the robot, especially the robot that may be able to deal with a comfortable mug of hot coffee, the thin handle may be a much less reliable absorption point than a more revival of the mug itself.

Debindnd’s approach to automatic safety

Vikas CentianDeepMind, the project’s automatic safety, says that the team has taken a layer of safety. It begins with classic material safety control tools that run things like avoiding collision and stability, but also include “semantic safety” systems that establish both instructions and the consequences of following them. Sindhwani, who “trained to evaluate whether it is a possible procedure in a specific scenario, says, says Sindhwani, who” trained to evaluate whether it is a possible procedure in a specific scenario, “says Sindhwani, who” trained to evaluate whether it is a possible procedure in a specific scenario, “says Sindhwani, says that these systems are the most sophisticated in the geminen of robotics, that these systems are they are The most advanced in the model of robots.

And since “safety is not a competitive endeavor,” says Sindhwani, DeepMind releases a new data set and what he calls ASIMOV IndexWhich aims to measure the ability of the model to understand the rules of proper life. The standard contains each of the questions about visual scenes and text scenarios, putting the opinions of models about things such as the mixing of bleaching and vinegar (mix chlorine gas) and putting a soft game on a hot stove. In the journalistic briefing, Sindhwani said that Gemini models have a “strong performance” on this standard, and Technical report It showed that the models got more than 80 percent of the correct questions.

DEPMIND partnerships

Again in December, DeepMind and Humanoid Robotics Apptronik A partnershipParada says that the two companies are working together to “build the next generation of human robots with Gemini in essence.” Deepmind also provides its models for a group of “trusted laboratories”: Greater robotsand Early movement robotsand Boston dynamicsAnd Charming tools.