In UC Berkeley, scientists in robotic AI and the Laboratory Sergey Levine looked at the table where the tower of 39 blocks stood. Then the white and black robot, its only limb doubled as a hosted giraffe, approached towards the tower and the black leather whip. Through what might seem to an occasional viewer as a miracle of physics, the whip hit exactly in the right place to send a single block flying from the magazine, while the rest of the tower remained structurally healthy.
This task, known as “Jenga Whipping”, is a hobby that people with dexority and reflexes chase. Now it has been mastered by robots, thanks to the novel, and the prepared training method. By learning from human demonstrations and feedback, as well as its own experiments in the real world, this training protocol teaches robots to perform complicated tasks such as a whipped -cream with 100% success. What’s more, robots are taught at impressive speed, allowing them to read within one to two hours, how to perfectly assemble to calculate the motherboard, build a shelf and more.
Laying AI was powered by the Learning robot field tried to break the challenge to teach mechanical activities that are unpredictable or complicated, unlike a single action, such as repeatedly picking up the object from a special place on the conisyors belt. To solve it, when Ary, Levine Laboratory went on what is called “strengthening learning”.
Postdoctor researcher Jianlan Luo explained that in the strengthening learning the robot will try to task in the real world and use the camera feedback from his mistakes to eventually manage this skill. When the team first announced new software using this approach in early 2024, Luo said they were felt that others could quickly replicate their success using open source code themselves.
This autumn, the research team Levine, Luo, Charles Xu, Zheyuan Hu and Jeffrey Wu released a technical report on its latest system, which has received the whipping of Jeng. This new and improved added in the version of human intervention. For a special mouse that controls the robot, one can correct the robot race and these repairs can be incorporated into the proverbial memory bank of the robot. Using the AI method called Strengthening Learning, the robot analyzes the sum of all its experiments – assisted and without assistance, successful and unsuccessful – to better perform its task. Luo said that one has to hit less, as the robot learned from experience. “I needed to watch the robot for perhaps the first 30% or something, and then I could gradually care,” he said.
Sign up today and save 40% We Conference Pass!
The laboratory carried out its robotic system through the gautlet complicated tasks outside Jung whipping. The robot turned the egg in the pan; hand over the object from one arm to another; And he assembled a motherboard because the dashboard and divorce belt. Scientists chose these challenges because they were diverse and, according to Lu’s words, “all the magic of uncertainty in performing robotic tasks in a complex real world”.
The task of the divorce belt performed in terms of difficult. Every time the robot interact with the divorce belt – imagine trying to manipulate a floppy necklace on two pegs – had to predict and react to this change.
Jenga whipping represents a different kind of challenge. Includes physics that is difficult to model, so it is effective to train the robot using separate simulations; The real world experience was critical.
Scientists also tested the adaptability of robots by introducing accidents. They would force the monster to open up, dropping an object or moving a Maherboard when the robot tried to install a microchip, training it to respond to a shift situation that could meet outside the laboratory environment.
At the end of the training, the robot could make them 100% of the time correctly. Scientists compared their results with the common method of “copying my behavior” known as cloning behavior that was trained on the same love for demonstration data; Their new system made robots faster and more accurate. These metrics are crucial, Luo said, because the bar for the robot connection is very high. Regular consumers and industrialists doe want to buy an unloaded robot. Luo stressed that, in particular, custom -made production process, such as those that are often used for electronics, cars and aviation parts, could benefit from robots that can connect and learn a number of tasks.
The first time the robot conquered the challenge of Jengha Whipping, “That really shocked me,” Luo said. “Yanga’s task is very difficult for most people.” I tried it with a whip in my hand; I had a success rate for 0%. “And even if it stacks against the adept of a human yck, he added that the robot is likely to overcome man because he has no muscles that will be elevated.
The new Levine Lab educational system is part of a wider trend in robotic innovations. Over the past two years, the larger field has shifted jumps and bonds powered by industry and AI, providing engineers to the turbocharged tool analysis or image inputs that the robot could observe. Professors and scientists Berkeley are part of this Upswell in innovation; Various top robotic companies that receive considerable funding or even public financing have links to the campus.
Levine co -founded Robotics Company Physical Intelligence (PI), which is currently worth $ 2 billion for its progress towards creating software that can work for different robots. In its last round of financing, Pi Investors received $ 400 million, including Jeff Bezos and Openi. In 2018, Professor Ken Goldberg and other scientists Berkeley, who received approximately $ 67 million, created robotics; The company creates robots trained through AI simulations that grab and sort land into various containers, which is necessary for business in electronic trading.
Pieter Abbeel, director of Berkeley Artificial Intelligence Research Lab, was created together by AI Robotics Startup Covariant, whose models and brain confidence were included in Amazon last year. And Homayoon Kazerooni, a professor of mechanical engineering, was founded by a publicly traded EKSO Bionics company, which produces robotic “exoskeletons” for use of people with limited mobility.
As for LUO research, he is excited to see where the team and other scientists can push it. The next step, he said, would be a preliminary traban of the system with basic handling handling capacities, which would eliminate the need to learn those from zero and instructors directly to acquire more complex skills. The laboratory is also a matter of being an open resource so that other scientists can use and build on it.
“The key goal of this project is to make technology as accessible and user -friendly as an iPhone,” Luo said. “I firmly believe that the more people can use it, the greater the impact we can do.”
Editor’s note: This article was re -published from UC Berkeley News.