Robots can learn new tasks and thank you for AI techniques. Mint

Cambridge, inside the Labotics laboratory of the Toyota Research Institute (TRI) at Massachusetts, is busy cooking a group of robots. There is nothing special about it; Robot chefs have been around for some time. But these robots are more efficient than most, flip the pancakes, cut the vegetables and make pizza easily. The difference is that instead of being laborlessly programmed to perform its tasks, the Cambridge robot has only been taught an original set of skills. Using the miracles of Artificial Intelligence (AI), they quickly became better on those skills so that they became more skillful.

Despite their extraordinary culinary abilities, these robots are not lucky for career in catering. Gill Pratt, the chief scientist at Toyota, says, “If you give the robot the confidence to work in the kitchen, it will also be confident of working in a factory or a person’s house.” Cooking includes a lot of complex functions, such as raising items and placing items, pouring fluid and mixing the material. All this creates an ideal training ground to experiment with a new method of using generic AI to train robots known as a kitchen.

Dissemination, already used to help the AI ​​model generate images, developed as a way to speed up robot training by TRAI and Robotist in Columbia University and Massachusetts Institute of Technology (MIT) Has gone. To explain how the spread works, the vice -president of Robotics Research and a professor in MIT uses the function of a specific kitchen, teaching a robot how to load a dishwasher, a Bar its fellow machines are done with their cooking.

Traditionally, the robot is programmed with rems of the computer code. It can be manually manufactured or the arms and hands of the robot can be taken from distance to repeat the necessary actions. In Cambridge, a robot was taught to provide feedback, equipped with camera eyes and touch sensors, to take dishes in a remote-control manner and stack them in dishwashers. This included about 100 such performances, each separate, to deal with different objects and how they should be stacked.

Nevertheless, 100 performances are also not enough to cover every event, which comes in spread. The process is a bit like learning how to separate it and build a gizmo and try to recreate it again. For the image generation, it involves adding random “noise” to a photo until it does not become unfamiliar and then reverses the process to learn the stages involved in creating a new, realistic image.

For robot training, AI uses tasks that have been randomly taught to generate potentially new movements, which are later refined in useful functions that can deal with the new environment. It may be how to raise a plate placed on an unusual angle or a strange shape bowl. The robot will try new tasks until it succeeds in his work. Using spread Dr. Tedrek says that it was possible to train the robot in a few hours to load a dishwasher, while a traditionally programming took a year or more time.

After spreading to work for various tasks, researchers are now trying to bring hundreds of such tasks together that they call a large behavior model (LBM). It will suit a large language model (LLM), which is used to power AI services such as Chatgpt. Based on the information, instead of generating answers to the questions, on which an LLM is trained, an LBM has sets of behavior that can be used to generate new behaviors. In its simplest form, it means that the skills involved in taking grocery items from supermarket shelves (which has learned to be used to select components in cars that make cars to learn how to do) Can

Once these new skills acquired can then be moved from one robot to another in a wireless manner, called “fleet learning”. This will also help in speeding up robot training. Over time, even basic training can be rapid and simplified. Instead of moving your organs remotely, the robot can just display someone how to do the job.

To pursue this work, TRAI, which is located in the Silicon Valley, recently with Boston Dynamics. Developing widely running robots is seen as one of the world leaders, Boston Dynamics is working on a mild and small version of Atlas, its Hawking Humanoids, which can run, can jump And even cartwheels can do. The new atlas will provide an agile robot, which the TRAI aims to equip with an LBM.

The idea is that, initially at least, these robots will be deployed in factories, most likely to make vehicles (both TRAI and Boston Dynamics are part of large car manufacturers: Toyota is Japan’s largest car manufacturer and In 2021, a large South Korean manufacturer, Hyundai, a large South Korean manufacturer, bought a majority stake in Boston Dynamics). Factories are a relatively structured environment in which automation is already widely used, making the introduction of AI-operated Humanoids easier. Humanoids are seen as the most efficient shape to use in man-made environment rather than widely wheel or tracked robots. The same is true in homes.

Eventually car factories can produce mass robots themselves, which will reduce prices and allow them to introduce them in other areas, such as helping the elderly and disabled people care. Elon Musk has a similar strategy that is planned for a similar strategy for optimus, a humanoid AI-powered robot developed by its electric-car company, Tesla. Mr. Musk has not given any details about the form of AI that is using Tesla.

All this may look for a future in which humans are no longer necessary in factories. But, Dr. Pratt says, as manufacturing becomes more flexible, and a large variety of products is created on the same line, will ever become more dependent on the human work force to manage factory changes and maintain robots. Many hands make light work.