Sunday, April 21, 2024

Google DeepMind has new rules to make sure AI robots behave when tidying your home

by gemnews
0 comment
Google DeepMind has written a constitution to make AI-powered robots safer.
Google DeepMind has written a constitution to make AI-powered robots safer.
  • Robots that behave in your home while doing the chores might arrive sooner than you think.
  • Google DeepMind announced a set of new rules this week that aim to make AI-powered robots safer.
  • They seek to give robots a better grasp of objects, as well as clearer task instructions.

Google is betting that AI-powered robots will be indispensable to our domestic lives someday. They could organize our homes and figure out meal prep for the week. Maybe they'll even cook those meals for us, too.

The search giant knows, however, that this future will only emerge if the robots behave.

That's why, on Thursday, its Google DeepMind team introduced a set of advances that aim to ensure AI-powered robots act as humans ideally would when going about their daily activities.

Tasks like chores are, as DeepMind notes, pretty straightforward for humans. Robots, however, need to have a "high-level understanding of the world" embedded to be able to perform them.

Its first new advance that aims to give robots this high-level understanding is called AutoRT. DeepMind researchers say this is "a system that harnesses the potential of large foundational models" to make robots smarter.

These models, according to DeepMind, collect a bunch of experiential training data that allow the robots to learn more on the fly about their environment.

They include the kinds of large language models (LLMs) that power generative AI tools like ChatGPT or Google's Gemini. Uses of an LLM in a robot include suggesting "a list of creative tasks that the robot could carry out, such as 'place the snack onto the countertop.'"

Visual language models are also included. These are used in conjunction with a camera to help robots identify objects around them.

If robots are let loose in the home or an office, this ability is important for safety reasons; novel situations may arise that require the robots to adapt, for instance, to new people that might be around them or changes in the location of objects.

The DeepMind team said the AutoRT system was evaluated over seven months in various real-world situations. They found that it "safely orchestrated as many as 20 robots simultaneously" and up to 52 unique robots.

That safety is apparently the result of a "robot constitution" written by the company. It was inspired by the "Three Laws of Robotics," a set of rules devised by sci-fi writer Isaac Asimov, which include the instruction: "A robot may not injure a human being."

In practice, this constitution is meant to give the LLMs of the robots "a set of safety-focused prompts to abide by when selecting tasks."

Two other advances from Google are about making robots much more efficient than they are today.

A new system called SARA-RT, short for self-adaptive robust attention for robotic transformers, aims to speed up the decision-making processes involved in the neural networks underlying AI transformer models.

"While transformers are powerful, they can be limited by computational demands that slow their decision-making," DeepMind wrote in a blog.

The other advance involves a model called RT-Trajectory. This is meant to give robots a clearer way to understand what physical actions to take following a set of instructions.

Videos in a training dataset are overlaid with 2D sketches of how a robot would perform a task, providing "low-level, practical visual hints" in the process.

While robot vacuums like Roombas are already popular, DeepMind's advancements may take us a step closer to a reality where AI bots tackle all our most hated chores in ways we've only imagined.

Read the original article on Business Insider

Source

You may also like

Leave a Comment

About Us

We’re a media company. We promise to tell you what’s new in the parts of modern life that matter. 

@2024 – All Right Reserved.