Unlocking the Potential of Robotics with DayDreamer: A Real-World Training Algorithm

DayDreamer is an algorithm that quickly teaches robots new behaviors.
It can take a long time to train robots to perform tasks in real life. This involves creating a simulator that is fast and efficient, running numerous tests on it and then translating the learned behaviors into the real world. The performance of robots in simulations is often not the same as the one achieved in real life due to unpredictability in the task or environment.

Researchers from the University of California, Berkeley have developed DayDreamer, a new tool that can be used to teach robots how to perform real-world tasks better. The researchers’ approach, described in a pre-published paper on arXiv is based upon learning models of the environment that allow robots predict the outcome of their actions and movements, reducing the necessity for extensive training in real-world.

TechXplore reported that Danijar hafner, a researcher who conducted the study, said, \”We wanted robots to learn continuously in the real-world, without creating a simulation environment.\” \”We only had experience with world models in video games, so we were super excited to learn that the same algorithm allowed robots to rapidly learn in the real-world, too!\”

Source:
https://techxplore.com/news/2022-07-daydreamer-algorithm-quickly-robots-behaviors.html