Incredible general robots show us future housework

Incredible general robots show us future housework

by

The emerging starting physical intelligence is not interested in building robots. Instead, the team remembers something better: powering hardware with continuous learning of the general “brain” of AI software, so existing machines will be able to autonomically perform growing tasks that require accurate movements and dexterity – including housework.

Over the past year, we have seen robotic dog dance, even some equipped to shoot flames and more and more advanced humanoids and machines built for specialized roles on assembly lines. But we’re still waiting for our Rosey robot Jetsons.

But we can be there early. The physical intelligence of San Francisco (PI) revealed its general AI model for robotics that can seize existing machines to perform various tasks – in this case was washed out of dryer and folding clothing, gently wrap the eggs in their container, grinding coffee and coffee beans and “Bussing “Tables. It is not a section to imagine that this system that is these mobile metal helpers who roll into the house, vacuum, pack and unpack the dishwasher, create a bed, look into the refrigerator and pantry to catalog your content and come to dinner – A ,, a ,, a, hey, why not, he also cooks the dinner.

It is with this vision that it will destruction its “general robotic basic model of the robot” known as π0 (Pi-zero).

“We believe this is the first step towards our long -term object of development artificial intelligence, so users can simply ask robots to do any task they want, as they can ask large language models (LLM) and chatbot assistants,” the company explains . “Like LLMS, our model is trained for wide and diverse data and can follow different text instructions. Unlike LLMS, it includes pictures, text and events, and gains physical intelligence by training embodied experiences from robots, learning to directly emit low levels at low levels of motor commands through new architecture.

In his research, PI-Vero shows how many jobs requiring different levels of dexterity and movements can be done by hardware of trained AI. A total of the basic model performed 20 tasks, all required different skills and manipulations.

“Our goal in choosing these tasks is not to solve any specific application, but to start providing our model of general understanding of physical interactions – the initial basis for physical intelligence,” the team notes.

Now I am the last person in the New Atlas who was enthusiastic about robotics, to a large extent what we were specialized by machines – and to be honest, I had my filling of humanoids moving boxes from point A to the point to B. In biology Specialists are very good to use one niche – for example bees, butterflies and koala – and do it exceptionally well. This means, up to external forces, such as loss of habitat or illness, reveals their limitation.

The generalist – like Racoon or Grizzly Bear – may not be as good in casting one niche as the eager, but are a much more adaptable range of habitats and food sources. Which eventually causes them to be more suitable for dynamic changes in the environment.

Similarly, municipal robots will be able to do more than professionally to build a brick wall; And, able to learn, they will be able to see various challenges in the physical world and continue your skills developing.

Pre-training model Vizi-Jazyka pi-zero on the Internet (VLM) compared to the flow for synchronizing its movements with its learning AI. Its pre -school included 10,000 hours of “skillful handling data” from seven different robot configurations and 68 tasks. In addition, these were existing data files manipulating robot manipulation from Ox, Droid and Bridge.

“Manipulation handling the robot handling requires a pi-zero to issue engine commands at a high frequency, up to 50 times per second,” the team notes. “To provide this level of dexterity, we have developed a new method for increasing pre -trained VLM with continuous action outputs through a flow comparison, a variant of broadcasting models. Starting with various robotic data and VLM pre -trained on internet data. We train our model comparing streams and comparing languages ​​of visions, which can then be a post-press on high quality robotic data to solve the anger of the tasks down.

“According to our knowledge, this is a large pre -school mix that is ever used for a robot manipulation model,” scientists noted in their study.

While the company is still in its early days of research and development, co-founder PI and CEO Karol Hausman-scientist who previously worked on Robotice in Google-Believes, its basic model will overcome existing obstacles to generalization, involves the love of time and costs associated with training Hardware on data on the physical world to learn new tasks. The PI team also included co -founder Sergey Levine, who has Pieneed Robotics Development at Stanford University and Brian Ichter, a scientist on Google.

In 2023, the satirist and architect Karl Sharro became a viral tweet: “People doing hard work on minimum wages, robots write poetry and color is not the future I wanted.” Of America went to strike and saw a bleak path in front of a creative face in the face of this new age technology.

And while AI may still come-a already come for many of our jobs (you don’t have to remind us of journalists), the vision Pi feels more in harmony with those of the mid-20 futurists. . Call me a naive, maybe, a goal, if a robot comes for my homework, it can take it.

You can see more videos about the exercise that the team made on the PI-Zero blog post, but here the one that will hit its impressive and delicate work.

https://www.youtube.com/watch?v=23ljudi2x4q

Our constelled eggs

Research work on the development and training of PI-zero can be found here.

Source: Physical Intelligence

Leave a Comment