For the Toyota research laboratory and its partners, AI in robotics is on the verge of the same kind of revolution as with generative AI. Explanations.
This will also interest you
[EN VIDÉO] Interview: are avatars the future of robotics? Science fiction often uses avatars, these robotic doubles controlled remotely. But are they…
Like the revolution triggered by the arrival of ChatGPT, here comes a moment that could be its equivalent for robotics. In this world, investments are increasing and are turning towards the creation of autonomous humanoid robots capable of taking over the tasks of workers. But to achieve this, it is necessary for them to learn each task and at the moment, it is quite laborious and limited.
However, Toyota, MIT and Columbia Engineering claim having developed a new approach to learning AI. It significantly accelerates robot training time. The robot used has two arms and to train it, the scientists taught it how to cook… Initially, the robot has to follow the subtleties of an operator’s movements, a bit as if a person’s hands were guided to show them a gesture. The operator who has a controller in each hand can manipulate the movements of the robot’s arms. He can teach him how to beat an egg with the right tool, for example. So that the operator can feel what the robot is touching and with what level of pressure it comes into contact with an object, the controls have haptic feedback. The operator can thus teach the AI how to gauge the pressure on an object’s grip.
These robotic arms aren’t exactly the stuff of dreams, but the AI that drives them could very well be integrated into real humanoid robots with fingers and capable of impeccably reproducing human tasks. Learning is done thanks to an operator who guides the robot’s movements so that it can carry out its task correctly. ©Toyota
The ChatGPT of robotics is on its way!
Once the AI has learned to carry out a task repeated several times, but with slightly different conditions, it will generate its own internal model based on an algorithm highlighting what is good or less good. It will then run thousands of simulations to perfect its task. And it’s quite fast, since the team of researchers explains that the teaching of the tasks takes place in the afternoon and that the AI will learn during the night. The next morning, she is able to reproduce the tasks perfectly independently. So far, the team has used this approach to quickly train the robots in more than 60 small tasks around the kitchen. But Toyota aims to achieve learning of more than 1,000 tasks by the end of 2024. Above all, the firm is developing what it believes to be the first large-scale behavioral model (LBM) for robotics , that is to say the equivalent of the large language model (LLM) specific to ChatGPT. The difference is that unlike AI which ingests billions of written data to achieve very good results, it is much less obvious for robotics. AI must be able to observe how a human performs these tasks in real life. It takes much longer and yet the AI developed by Toyota shows that it is possible to gain considerably in speed.
Never miss a single Futura magazine by subscribing! Enjoy the comfort of receiving your magazine directly in your mailbox, and at a preferential rate.
By choosing our 1-year subscription offer, you will receive the next 4 issues of Mag’ Futura (148 pages to decipher the major challenges of today and tomorrow) for 1 year at only €4/month.
Futura is an independent and committed scientific media which needs its readers to continue to inform, analyze and decipher. To encourage this approach and discover our next publications, subscription remains the best way to support us.
#MIT #Toyota #launch #artificial #intelligence #revolution #robotics #fastlearning #humanoids